PaperlayerView exact revision on arXiv

Source-linked AI summary

The Neuro-Symbolic Inverse Planning Engine (NIPE): Modeling Probabilistic Social Inferences from Linguistic Inputs

Lance Ying, Katherine M. Collins, Megan Wei, Cedegao E. Zhang, Tan Zhi-Xuan, Adrian Weller, Joshua B. Tenenbaum, Lionel Wong

arXiv:2306.14325v2cs.AIcs.LG

TL;DR

Prior inverse-planning models face challenges representing complex worlds from language, and direct LLM reasoning has been evaluated on restricted tasks. NIPE translates linguistic scenarios into symbolic probabilistic models for Bayesian goal inference; it correlates strongly with human judgments and performs more consistently across task variants than direct LLM baselines.

  • Problem

    Prior inverse-planning models largely rely on images or demonstrations and often require world models specified in advance or learned through costly exploration, while direct LLM evaluations have used restricted tasks.

    Source: §1 · p. 2

  • Method

    NIPE uses an LLM to translate language into program expressions conditioning a probabilistic model over agents and environments, then applies Bayesian inverse planning to infer goals.

    Source: §1 · p. 3, §3.3 · p. 6

  • Results

    NIPE-human correlation was R=0.927 (95% CI=(0.877, 0.953)), versus GPT-4 Rating R=0.658 (CI=(0.403, 0.79)) and GPT-3.5 Rating R=0.100 (CI=(-0.145, 0.361)).

    Source: §5 · pp. 6–7

  • Takeaways & Limitations

    NIPE correlates with human judgments across the task variants, while LLM baselines are less robust to complex world dynamics and spatial information.

    Source: §5 · p. 7

    The LLMs were evaluated with simple prompting, so whether results persist or improve with other prompting and evaluation methods remains open.

    Source: §6 · p. 8

Abstract

from arXiv · show

Human beings are social creatures. We routinely reason about other agents, and a crucial component of this social reasoning is inferring people's goals as we learn about their actions. In many settings, we can perform intuitive but reliable goal inference from language descriptions of agents, actions, and the background environments. In this paper, we study this process of language driving and influencing social reasoning in a probabilistic goal inference domain. We propose a neuro-symbolic model that carries out goal inference from linguistic inputs of agent scenarios. The "neuro" part is a large language model (LLM) that translates language descriptions to code representations, and the "symbolic" part is a Bayesian inverse planning engine. To test our model, we design and run a human experiment on a linguistic goal inference task. Our model closely matches human response patterns and better predicts human judgements than using an LLM alone.

1. Introduction

The paper studies how language about agents, actions, and environments informs goal inference. NIPE translates linguistic scenarios into a probabilistic model for Bayesian inverse planning, and the authors report that it predicts human judgments better than direct LLM baselines.

1. Introduction

  • Language can describe actions, mental states, abstract world conditions, and questions, and information about one aspect can shape predictions about what agents know, want, or do.Source: §1 · p. 1
  • Prior inverse-planning models largely infer from images or demonstrations, while requiring world models to be specified in advance or learned through costly exploration.Inferring mental states or even framing the task can also be difficult from images and demonstrations alone.Source: §1 · pp. 1–2
  • Direct LLM social-reasoning evaluations have often used restricted tasks, while robust planning remains an outstanding challenge for LLMs.The paper focuses on language describing worlds and actions to infer the unseen goals motivating those actions.Source: §1 · pp. 2–3
  • NIPE translates language into program expressions that condition a probabilistic model over agents and environments, then performs Bayesian inverse planning to infer goals.The model builds on rational meaning construction for language-informed cognition.Source: §1 · p. 3
  • The evaluation compares NIPE with human judgments on linguistic inverse-planning tasks designed to jointly involve agent actions and changing world information.Source: §1 · p. 3
  • NIPE strongly predicts human judgments across varied scenarios and outperforms LLM baselines on the goal-inference tasks.GPT-4 correlates moderately with some judgments, particularly on simpler scenarios.Source: §1 · p. 3

2. Linguistic inverse planning tasks and human experiment

The paper evaluates linguistic goal inference in a gameshow obstacle-course domain, varying the information provided about environments, rules, and agent actions. Human participants judged likely goals across these scenarios.

2. Linguistic inverse planning tasks and human experiment

  • The task set varies information about world configuration, world dynamics, and observed actions to test how people and models infer goals in varied environments.Source: §2 · p. 3

2.1. Planning and inverse planning concepts in the Gameshow Domain

  • Players navigate an obstacle course toward one of several trophies, with locked doors and keys requiring multi-step reasoning about longer-term plans.Each scenario description is paired with a query asking which trophy the player is trying to reach.Source: §2.1 · p. 3

2.2. Task variants

  • The 18 stimuli span four variants: generic keys, same-colored keys, different-colored keys, and spatial navigation without keys or doors.The color variants specify whether keys must match or differ from door colors; spatial scenarios describe layouts and actions.Source: §2.2 · p. 3

2.3. Human experiment

  • Forty-six participants each judged seven stimuli from one variant on a 1–7 goal-likelihood scale, with 14–16 participants seeing each scenario.Six participants were excluded because fewer than 30% of their most-likely-goal judgments agreed with consensus.Source: §2.3 · p. 4

3. Our Model: The Neuro-Symbolic Inverse Planning Engine (NIPE)

NIPE converts linguistic descriptions into a symbolic probabilistic model over agents and environments, then uses Bayesian inverse planning to infer goals. Its components address language translation, model construction, and inference over observed actions.

3. Our Model: The Neuro-Symbolic Inverse Planning Engine (NIPE)

  • NIPE decomposes linguistic goal inference into language-to-code translation, construction of a probabilistic model over agents and environments, and Bayesian inverse planning.The framework uses explicit planners and principled probabilistic inference for complex models.Source: §3 · pp. 4–5

3.1. Language-to-code translation using LLMs

  • A few-shot LLM translates sentences into code that composes into a symbolic model across a multi-sentence scenario.The implementation uses GPT-3.5 and rejects translations that fail syntactic or semantic checks, including validity and executability.Source: §3.1 · p. 5

3.2. Constructing a probabilistic generative model over agents and environments from language

  • The translated language conditions distributions over possible worlds and agent actions, including map layouts, domain rules, and descriptions of subgoals or action sequences.Spatial conditions constrain sampled maps; PDDL rules specify action preconditions and effects, while observed behavior can condition subgoals or low-level actions.Source: §3.2 · pp. 5–6
  • Scene representations and domain rules form a concrete PDDL planning domain nested within a Bayesian model of agents acting in that environment.The PDDL domain can be transferred to another setup by changing operator definitions or selecting another predefined domain.Source: §3.2 · p. 5

3.3. Goal Inference through Bayesian Inverse Planning

  • Bayesian inverse planning combines a goal prior with the likelihood of observed actions under each goal, then normalizes the joint probabilities into a posterior over goals.The prior favors goals closer by shortest-plan cost; a Boltzmann-rational policy models action choice, and A* computes optimal-plan costs. The model does not account for forms of approximate rationality.Source: §3.3 · p. 6

4. Model Experiments

The experiments compare NIPE with human judgments and two direct-LLM baselines on the goal-inference stimuli. NIPE uses a single valid sampled representation per stimulus, while the LLM baselines answer the same scenarios directly.

4. Model Experiments

  • The experiments compare NIPE with human judgments and two LLM baselines to assess the value of explicit Bayesian inverse planning.Source: §4 · p. 6
  • For each stimulus, NIPE translates the scenario into symbolic conditions and computes a posterior over trophies using one valid sample of PDDL rules and spatial environments.Pilot experiments found limited variability across LLM and generative-model samples, and goal inference can be performed without sampling.Source: §4 · p. 6
  • The GPT-3.5 and GPT-4 baselines receive zero-shot prompts containing the human instructions, scenario, and query, with 30 answers sampled per stimulus.Answers that do not match the required format are rejected.Source: §4 · p. 6

5. Results and discussion

NIPE’s goal probabilities track human judgments strongly overall and across all four task variants, while direct LLM judgments are weaker, especially on complex variants. The task-level results show that stated world rules and efficient spatial planning expose differences between model judgments and human reasoning.

5. Results and discussion

  • R=0.927 (95% CI=(0.877, 0.953)) for NIPE, compared with GPT-4’s R=0.658 (CI=(0.403, 0.79)) and GPT-3.5’s R=0.100 (CI=(-0.145, 0.361)) overall.Correlations compare model goal probabilities with mean human judgments across stimuli.Source: §5 · pp. 6–7, Figure 2 · p. 7
  • NIPE correlates strongly with human judgments across all four variants, whereas LLM baselines generally fare worse on the Color-Different and Spatial variants.Color-Different adds a counterintuitive world rule; Spatial requires reasoning about an efficient path-planning agent.Source: §5 · p. 7, Figure 3 · p. 8
  • GPT-3.5 correlates negatively on Color-Different and nearly not at all on Spatial; GPT-4 is moderate on Spatial but nearly uncorrelated on Color-Different.The results suggest GPT-4 handles the key-color rule unevenly, while its Spatial performance may reflect distance heuristics rather than robust path planning.Source: §5 · p. 7
  • In a Spatial example, humans and NIPE favor the bronze trophy after Alice walks past the silver one, while both LLM baselines favor the nearer silver trophy.The human judgment is consistent with treating Alice as an efficient actor whose goal lies ahead of her path.Source: §5 · p. 7

6. Conclusion and Future Directions

The authors identify limits in the prompting and translation choices and outline extensions to the task domain. Future work includes testing alternative prompts and broader social reasoning scenarios, and exploring a more fully Bayesian treatment of language translation.

6. Conclusion and Future Directions

  • The LLM baselines use a simple prompting regime, so future evaluations should test other methods, including chain-of-thought, and a wider diversity of prompts to control potential surface-cue biases.The authors note that prompts might bias models toward surface associations, such as gold being more valuable than silver.Source: §6 · p. 8
  • NIPE uses Bayesian inference for social reasoning, but its LLM-based language translation is bottom-up and non-Bayesian.The authors propose exploring a more fully Bayesian approach that models language generation from symbolic task representations and performs inference directly from language.Source: §6 · p. 8
  • The gameshow domain could be extended to belief inference, boundedly rational agents, collaborating or competing agents, and multimodal stimuli.The authors also propose increasing difficulty by expanding the objects, actions, and planning steps.Source: §6 · p. 8

A.2. Human linguistic inverse planning experiment

Participants rated how likely an agent was to be trying to reach each possible trophy on a 1–7 scale. A mismatch between the Generic and Different-Colored Keys conditions was corrected partway through the study.

A.2. Human linguistic inverse planning experiment

  • Participants rated the likelihood that the agent was trying to reach each possible trophy on a 1–7 scale.The question asked participants to consider each trophy separately.Source: §A.2 · p. 11
  • The supplied scale labels 1 as definitely not, 3 as somewhat unlikely, 4 as equally likely or unlikely, and 7 as definitely the trophy.Source: §A.2 · p. 11
  • Some participants in the Generic condition saw a Different-Colored Keys stimulus, and vice versa; the mismatch was corrected partway through the study.Source: §A.2 · p. 11

A.3. LLM Baselines Experiment

The LLM baseline used a fixed GPT-4 checkpoint and sampled responses for each stimulus, rejecting outputs that did not meet the required format.

A.3. LLM Baselines Experiment

  • The baseline used gpt-4-0314, a fixed, non-updatable checkpoint from March 14.The passage gives maximal reproducibility as the reason for using the fixed checkpoint.Source: §A.3 · p. 11
  • For each stimulus, the researchers sampled until they obtained 20 well-formed responses.Source: §A.3 · p. 11
  • Samples in an incorrect format were rejected.Source: §A.3 · p. 11

Example LLM Prompt for World Configuration

The example prompt describes trophies behind colored doors and Alice picking up a red key. The LLM output represents the scenario as structured fields, including goals, locations, obstacles, keys, and an observation.

Example LLM Prompt for World Configuration

Example LLM Prompt for PDDL Domain Definition

The examples show how the LLM is prompted to represent a domain and how GPT-4’s baseline response can conflict with the Different-Colored Keys rule. The sample stimuli also illustrate variations in key rules, trophy routes, and spatial descriptions.

Example LLM Prompt for PDDL Domain Definition

A.4. An example of LLM baseline result

  • In the Different-Colored Keys example, GPT-4 rates gold and silver 6 and bronze 1, while humans and NIPE favor bronze and give gold and silver roughly equal low likelihoods.The authors attribute the discrepancy to GPT-4 judging gold and silver highly likely despite the stated rule that yellow keys cannot open yellow doors.Source: §A.4 · p. 13

A.5. Sample stimuli for each experimental condition

  • Another stimulus makes gold require two doors, silver one door, and bronze no door, while a separate layout puts gold and silver behind a yellow door reached through green and bronze behind red.Source: §A.5 · p. 13
  • The Color-same sample specifies one lock per door, keys that work on all locks but only once, and Alice picking up two keys.Source: §A.5 · p. 13
  • One Different-Colored Keys stimulus places rooms behind yellow-and-green, red, and yellow locks, then has Alice pick up a yellow key under the different-color rule.Source: §A.5 · p. 14
  • A spatial stimulus places trophies in rooms A, B, and C and describes Alice walking three steps east, with gold one step east, silver three steps south, and bronze three steps east of silver.Source: §A.5 · p. 14

Related papers

Loading 2306.14325v2…