Drawing real world social inferences usually requires taking into account information from multiple modalities. Language is a particularly powerful source of information in social settings, especially in novel situations where language can provide both abstract information about the environment dynamics and concrete specifics about an agent that cannot be easily visually observed. In this paper, we propose Language-Informed Rational Agent Synthesis (LIRAS), a framework for drawing context-specific social inferences that integrate linguistic and visual inputs. LIRAS frames multimodal social reasoning as a process of constructing structured but situation-specific agent and environment representations – leveraging multimodal language models to parse language and visual inputs into unified symbolic representations, over which a Bayesian inverse planning engine can be run to produce granular probabilistic judgments. On a range of existing and new social reasoning tasks derived from cognitive science experiments, we find that our model (instantiated with a comparatively lightweight VLM) outperforms ablations and state-of-the-art models in capturing human judgments across all domains.

Language-Informed Synthesis of Rational Agent Models for Grounded Theory-of-Mind Reasoning On-the-fly

The relationship between the language a speaker conveys and the meaning a listener recovers is often probabilistic and sensitive to context. Numerous strategies attempt to estimate such a mapping, often leveraging recursive Bayesian models of communication. In parallel, large language models (LLMs) have been increasingly applied to semantic parsing applications, tasked with inferring logical representations from NL. While existing LLM explorations have been largely restricted to literal language use, in this work, we evaluate the capacity of LLMs to make inferences over pragmatic utterances. Specifically, we explore the case of threshold estimation on gradable adjectives, e.g., “strong”, contextually conditioned on a strength prior, then extending to composition with qualification, negation, polarity inversion, and class comparison. We find that LLMs can derive context-grounded, human-like inferences over interpretations for complex utterances such as “very strong for a beginner”, but struggle with composing such amortized inferences with negation or polarity inversion.

LLMs are pretty strong pragmatic reasoners: Investigating amortized inferences about gradable adjectives in large language models

The relationship between communicated language and intended meaning is often probabilistic and sensitive to context. Numerous strategies attempt to estimate such a mapping, often leveraging recursive Bayesian models of communication. In parallel, large language models (LLMs) have been increasingly applied to semantic parsing applications, tasked with inferring logical representations from natural language. While existing LLM explorations have been largely restricted to literal language use, in this work, we evaluate the capacity of LLMs to infer the meanings of pragmatic utterances. Specifically, we explore the case of threshold estimation on the gradable adjective "strong", contextually conditioned on a strength prior, then extended to composition with qualification, negation, polarity inversion, and class comparison. We find that LLMs can derive context-grounded, human-like distributions over the interpretations of several complex pragmatic utterances, yet struggle composing with negation. These results inform the inferential capacity of statistical language models, and their use in pragmatic and semantic parsing applications.

Evaluating statistical language models as pragmatic reasoners

Information Extraction (IE) for semi-structured document images is often approached as a sequence tagging problem by classifying each recognized input token into one of the IOB (Inside, Outside, and Beginning) categories. However, such problem setup has two inherent limitations that (1) it cannot easily handle complex spatial relationships and (2) it is not suitable for highly structured information, which are nevertheless frequently observed in real-world document images. To tackle these issues, we first formulate the IE task as spatial dependency parsing problem that focuses on the relationship among text tokens in the documents. Under this setup, we then propose SPADE (SPAtial DEpendency parser) that models highly complex spatial relationships and an arbitrary number of information layers in the documents in an end-to-end manner. We evaluate it on various kinds of documents such as receipts, name cards, forms, and invoices, and show that it achieves a similar or better performance compared to strong baselines including BERT-based IOB taggger.

Leveraging Language to Learn Program Abstractions and Search Heuristics

Core knowledge concepts such as object behavior principles provide a rich inventory of primitives for thinking and learning in the natural world. However, it remains unexplored how these concepts are reused for problem-solving and communication in highly abstract domains.

We analyze a large-scale natural language study drawing on the Abstraction and Reasoning Corpus (ARC), a set of highly abstract visual tasks where solvers construct outputs from input grids according to an inferred pattern. ARC explicitly incorporates core knowledge principles without any real-world objects. In the study, subjects solved and communicated the inferred patterns of ARC tasks via written explanations for other subjects attempting to solve tasks using only the explanations.

We examine how subjects solve, communicate, and interpret these explanations, and we show that subjects use fundamentally abstract core knowledge properties–object cohesion and contact causality–to reason about, understand, and communicate the inference tasks with language.

Core knowledge objects in reasoning and language use for highly abstract inductive tasks

The Abstraction and Reasoning Challenge (ARC) is a set of tasks where one must induce a program from a few given input-output examples, and apply it to a new input. Although humans can easily solve most tasks, ARC is challenging for state-of-the-art algorithms. We hypothesize that humans use intuitive program induction, and interpret ARC tasks by constructing "natural programs'' in language.

We experimentally study "natural programs'' by formulating a two-player game: A participant solves and then communicates the program to another participant using natural language; the second participant must solve the task using the description alone. We find at least 361 out of 400 tasks can be solved from a natural language description, demonstrating that natural language is sufficient in transmitting these natural programs. 

We compared the natural language programs to computer programs constructed using two separate state-of-the-art program synthesis approaches, and conducted a study on leveraging natural language annotations to improve performance of program-synthesis tools.

Language annotated Abstraction and Reasoning Corpus

People think and learn abstractly and compositionally. These two key properties of human cognition are shared with natural language: we use a finite, composable vocabulary of nameable concepts to generate and understand a combinatorially large space of new sentence. In this paper, we present a domain of compositional reasoning tasks and an artificial language learning paradigm designed to probe the role language plays in bootstrapping learning. We discuss results from a language-guided program learning model suggesting that language can play an important role in bootstrapping learning by providing an important signal for search on individual problems, and a cue towards named, reusable abstractions across the domain as a whole. We evaluate adults on the same domain, comparing learning performance between those tasked with jointly learning language and solving reasoning tasks, and those who only approach the domain as a collection of inductive reasoning problems. We find that adults provided with abstract language prompts are better equipped to generalize and compose concepts learned across a domain than adults solving the same problems using reasoning alone.

Lionel Wong

7

78

SHORT BIO

OTHER AFFILIATIONS

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES