$\texttt{DIAMONDs}$: A Dataset for $\mathbb{D}$ynamic $\mathbb{I}$nformation $\mathbb{A}$nd $\mathbb{M}$ental modeling $\mathbb{O}$f $\mathbb{N}$umeric $\mathbb{D}$iscussions

Understanding multiparty conversations demands robust Theory of Mind (ToM) capabilities, including the ability to track dynamic information, manage knowledge asymmetries, and distinguish relevant information across extended exchanges. To advance ToM evaluation in such settings, we present a carefully designed scalable methodology for generating high-quality benchmark conversation-question pairs with these characteristics. Using this methodology, we create $\texttt{DIAMONDs}$, a new conversational QA dataset covering common business, financial or other group interactions. In these goal-oriented conversations, participants often have to track certain numerical quantities (say $\textit{expected profit}$) of interest that can be derived from other variable quantities (like $\textit{marketing expenses, expected sales, salary}$, etc.), whose values also change over the course of the conversation. $\texttt{DIAMONDs}$ questions pose simple numerical reasoning problems over such quantities of interest (e.g., $\textit{funds required for charity events, expected company profit next quarter}$, etc.) in the context of the information exchanged in conversations. This allows for precisely evaluating ToM capabilities for carefully tracking and reasoning over participants' knowledge states. Our evaluation of state-of-the-art language models reveals significant challenges in handling participant-centric reasoning, specifically in situations where participants have false beliefs. Models also struggle with conversations containing distractors and show limited ability to identify scenarios with insufficient information. These findings highlight current models' ToM limitations in handling real-world multi-party conversations.

Does RoBERTa Perform Better than BERT in Continual Learning: An Attention Sink Perspective

Continual learning (CL) aims to train models that can sequentially learn new tasks without forgetting previous tasks' knowledge. Although previous works observed that pre-training can benefit CL, it remains unclear whether a pre-trained model with higher downstream capacity also performs better in CL. In this paper, we observe that pre-trained models may allocate high attention scores to some 'sink' tokens, such as [SEP] tokens, which are ubiquitous across various tasks. Such attention sinks may lead to models' over-smoothing in single-task learning and interference in sequential tasks' learning, which may compromise the models' CL performance despite their high pre-trained capabilities. To reduce these effects, we propose a pre-scaling mechanism that encourages attention diversity across all tokens. Specifically, it first scales the task's attention to the non-sink tokens in a probing stage, and then fine-tunes the model with scaling. Experiments show that pre-scaling yields substantial improvements in CL without experience replay, or progressively storing parameters from previous tasks.

Look Hear: Gaze Prediction for Speech-directed Human Attention

For computer systems to effectively interact with humans using spoken language, they need to understand how the words being generated affect the users' moment-by-moment attention. Our study focuses on the incremental prediction of attention as a person is seeing an image and hearing a referring expression defining the object in the scene that should be fixated by gaze. To predict the gaze scanpaths in this incremental object referral task, we developed the Attention in Referral Transformer model or ART, which predicts the human fixations spurred by each word in a referring expression. ART uses a multimodal transformer encoder to jointly learn gaze behavior and its underlying grounding tasks, and an autoregressive transformer decoder to predict, for each word, a variable number of fixations based on fixation history. To train ART, we created RefCOCO-Gaze, a large-scale dataset of 19,738 human gaze scanpaths, corresponding to 2,094 unique image-expression pairs, from 220 participants performing our referral task. In our quantitative and qualitative analyses, ART not only outperforms existing methods in scanpath prediction, but also appears to capture several human attention patterns, such as waiting, scanning, and verification.

AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents

Autonomous agents that address day-to-day digital tasks (e.g., ordering groceries for a household), must not only operate multiple apps (e.g., notes, messaging, shopping app) via APIs, but also generate rich code with complex control flow in an iterative manner based on their interaction with the environment. However, existing benchmarks for tool use are inadequate, as they only cover tasks that require a simple sequence of API calls. To remedy this gap, we built $\textbf{AppWorld Engine}$, a high-quality execution environment (60K lines of code) of 9 day-to-day apps operable via 457 APIs and populated with realistic digital activities simulating the lives of ~100 fictitious users. We then created $\textbf{AppWorld Benchmark}$ (40K lines of code), a suite of 750 natural, diverse, and challenging autonomous agent tasks requiring rich and interactive code generation. It supports robust programmatic evaluation with state-based unit tests, allowing for different ways of completing a task while also checking for unexpected changes, i.e., collateral damage. The state-of-the-art LLM, GPT-4o, solves only ~49% of our 'normal' tasks and ~30% of 'challenge' tasks, while other models solve at least 16% fewer. This highlights the benchmark's difficulty and AppWorld's potential to push the frontiers of interactive coding agents. The project website is available at https://appworld.dev/.

On Initializing Transformers with Pre-trained Embeddings

It has become common practice now to use random initialization schemes, rather than the pre-trained embeddings, when training transformer based models from scratch. Indeed, we find that pre-trained word embeddings from GloVe, and some sub-word embeddings extracted from language models such as T5 and mT5 fare much worse compared to random initialization. This is counter-intuitive given the well-known representational and transfer-learning advantages of pre-training. Interestingly, we also find that BERT and mBERT embeddings fare better than random initialization, showing the advantages of pre-trained representations. In this work, we posit two potential factors that contribute to these mixed results: the model sensitivity to parameter distribution and the embedding interactions with position encodings. We observe that pre-trained GloVe, T5, and mT5 embeddings have a wider distribution of values. As argued in the initialization studies, such large value initializations can lead to poor training because of saturated outputs. Further, the larger embedding values can, in effect, absorb the smaller position encoding values when added together, thus losing position information. Standardizing the pre-trained embeddings to a narrow range (e.g. as prescribed by Xavier) leads to substantial gains for Glove, T5, and mT5 embeddings. On the other hand, BERT pre-trained embeddings, while larger, are still relatively closer to Xavier initialization range which may allow it to effectively transfer the pre-trained knowledge.

CaT-BENCH: Benchmarking Language Model Understanding of Causal and Temporal Dependencies in Plans

Understanding the abilities of LLMs to reason about natural language plans, such as instructional text and recipes, is critical to reliably using them in decision-making systems. A fundamental aspect of plans is the temporal order in which their steps needs to be executed, which reflects the underlying causal dependencies between them. We introduce CaT-Bench, a benchmark of Step Order Prediction questions, which test whether a step must necessarily occur before or after another in cooking recipe plans. We use this to evaluate how well frontier LLMs understand causal and temporal dependencies. We find that SOTA LLMs are underwhelming (best zero-shot is only 0.59 in F1), and are biased towards predicting dependence more often, perhaps relying on temporal order of steps as a heuristic. While prompting for explanations and using few-shot examples improve performance, the best F1 result is only 0.73. Further, human evaluation of explanations along with answer correctness show that, on average, humans do not agree with model reasoning. Surprisingly, we also find that explaining after answering leads to better performance than normal chain-of-thought prompting, and LLM answers are not consistent across questions about the same step pairs. Overall, results show that LLMs' ability to detect dependence between steps has significant room for improvement.

Comparing Human-Centered Language Modeling: Is it Better to Model Groups, Individual Traits, or Both?

Natural language processing has made progress in incorporating human context into its models, but whether it is more effective to use group-wise attributes (e.g., over-45-year-olds) or model individuals remains open. Group attributes are technically easier but coarse: not all 45-year-olds write the same way. In contrast, modeling individuals captures the complexity of each person's identity. It allows for a more personalized representation, but we may have to model an infinite number of users and require data that may be impossible to get. We compare modeling human context via group attributes, individual users, and combined approaches. Combining group and individual features significantly benefits user-level regression tasks like age estimation or personality assessment from a user's documents. Modeling individual users significantly improves the performance of single document-level classification tasks like stance and topic detection. We also find that individual-user modeling does well even without user's historical data.

Modeling Complex Event Scenarios via Simple Entity-focused Questions

Event scenarios are often complex and involve multiple event sequences connected through different entity participants. Exploring such complex scenarios requires an ability to branch through different sequences, something that is difficult to achieve with standard event language modeling. To address this, we propose a question-guided generation framework that models events in complex scenarios as answers to questions about participants. At any step in the generation process, the framework uses the previously generated events as context, but generates the next event as an answer to one of three questions: what else a participant did, what else happened to a participant, or what else happened. The participants and the questions themselves can be sampled or be provided as input from a user, allowing for controllable exploration. Our empirical evaluation shows that this question-guided generation provides better coverage of participants, diverse events within a domain, comparable perplexities for modeling event sequences, and more effective control for interactive schema generation.

Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge-Intensive Multi-Step Questions

Recent work has shown that large language models are capable of generating natural language reasoning steps or Chains-of-Thoughts (CoT) to answer a multi-step question when prompted to do so. This is insufficient, however, when the necessary knowledge is not available or up-to-date within a model's parameters. A straightforward approach to address this is to retrieve text from an external knowledge source using the question as a query and prepend it as context to the model's input. This, however, is also insufficient for multi-step QA where \textit{what to retrieve} depends on \textit{what has already been derived}. To address this issue we propose IRCoT, a new approach that interleaves retrieval with CoT for multi-step QA, guiding the retrieval with CoT and in turn using retrieved results to improve CoT. Our experiments with GPT3 show substantial improvements in retrieval (up to 22 points) and downstream QA (up to 16 points) over the baselines on four datasets: HotpotQA, 2WikiMultihopQA, MuSiQue, and IIRC. Notably, our method also works well for much smaller models such as T5-Flan-large (0.7B) without any additional training.

POQue: Asking Participant-specific Outcome Questions for a Deeper Understanding of Complex Events

Knowledge about outcomes is critical for complex event understanding but is hard to acquire. We show that by pre-identifying a participant in a complex event, crowd workers are able to (1) infer the collective impact of salient events that make up the situation, (2) annotate the volitional engagement of participants in causing the situation, and (3) ground the outcome of the situation in state changes of the participants. By creating a multi-step interface and a careful quality control strategy, we collect a high quality annotated dataset of 8K short newswire narratives and ROCStories with high inter-annotator agreement (0.74-0.96 weighted Fleiss Kappa). Our dataset, POQue (Participant Outcome Questions), enables the exploration and development of models that address multiple aspects of semantic understanding. Experimentally, we show that current language models lag behind human performance in subtle ways through our task formulations that target abstract and specific comprehension of a complex event, its outcome, and a participant's influence over the event culmination.