OpenNovelty.org

Language and Experience: A Computational Model of Social Learning in Complex Tasks

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.4Download Report PDF
cognitive science; social learning; cultural learning; causal learning; bayesian models of cognition
Abstract:

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models human social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models—revealing how structured, language-compatible representations might enable human-machine collaborative learning.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper presents a computational framework modeling human social learning through joint probabilistic inference over structured world models, integrating linguistic advice with sensorimotor experience. It resides in the 'Computational Models of Social Learning' leaf under 'Integrating Language and Experience', which contains only two papers total. This sparse positioning suggests the work addresses a relatively underexplored niche within the broader taxonomy of 47 papers. The sibling paper focuses on integrating natural language hints with reward signals, indicating the leaf targets explicit language-experience fusion rather than purely instruction-following or purely experiential learning approaches.

The taxonomy reveals that most related work clusters in adjacent branches: 'Language-Guided Agent Learning and Planning' contains systems using instructions for task specification and hierarchical planning, while 'Experience-Based Learning and Adaptation' emphasizes memory retrieval and self-evolution mechanisms. The 'Bayesian Integration of Language and Reward' leaf sits nearby, focusing on probabilistic bandit models rather than structured world models. The paper's emphasis on social learning—modeling how humans share and interpret advice—distinguishes it from these neighboring directions, which typically address single-agent instruction following or reward-based adaptation without explicit social communication modeling.

Among 24 candidates examined across three contributions, the literature search found limited prior overlap. The core framework contribution (10 candidates examined, 0 refutable) and the language-accelerated inference contribution (10 candidates, 0 refutable) appear relatively novel within this search scope. However, the speaker model contribution (4 candidates examined, 1 refutable) shows more substantial prior work, suggesting that using language models to generate or interpret advice has precedent. The analysis explicitly notes this is a limited top-K semantic search, not exhaustive coverage, meaning the novelty assessment reflects only the examined subset of the field.

Given the sparse taxonomy leaf and limited refutation signals across most contributions, the work appears to occupy a distinct position within the examined literature. The social learning framing and structured world model approach differentiate it from instruction-following systems and bandit-based integration methods. However, the restricted search scope (24 candidates) and the presence of at least one overlapping prior work for the speaker model component suggest caution in claiming comprehensive novelty without broader literature coverage.

Taxonomy

Core-task Taxonomy Papers
47
3
Claimed Contributions
24
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: integrating linguistic guidance and direct experience in sequential decision-making. The field structure reflects a broad effort to combine symbolic instruction with trial-and-error learning across diverse problem settings. The taxonomy organizes work into several main branches: Language-Guided Agent Learning and Planning focuses on how natural language instructions shape policy search and task execution (e.g., Describe Explain Plan[3], Natural Language Task[4]); Experience-Based Learning and Adaptation emphasizes reinforcement learning, imitation, and adaptive mechanisms that refine behavior through interaction; Integrating Language and Experience addresses computational models that explicitly merge these two modalities, including social learning frameworks; Explanation and Feedback Mechanisms explores how agents generate or consume human-readable rationales and corrective signals (e.g., Reflexion[2], Policy Explanations Usability[1]); Evaluation and Benchmarking provides testbeds and metrics; Domain-Specific Applications targets robotics, healthcare, and other specialized contexts (e.g., EndoAgent[6], MobileSteward[7]); and Theoretical Foundations and Formal Models underpins the field with decision-theoretic and linguistic formalisms (e.g., Linguistic Decision Analysis[32], Sequential Linguistic Assessment[33]). A particularly active line of work examines how agents can leverage verbal feedback or self-generated explanations to accelerate learning, contrasting purely reward-driven methods with those that incorporate structured language (Reflexion[2], Human Discrete Feedback[22]). Another theme concerns the trade-offs between pre-specified linguistic knowledge and emergent strategies learned from environmental interaction, with some studies exploring hybrid architectures that dynamically weight instruction versus experience. The original paper, Language Experience Computational[0], sits within the Computational Models of Social Learning cluster under Integrating Language and Experience, closely aligned with Integrating Natural Language[8]. Its emphasis on social learning mechanisms distinguishes it from purely individual agent frameworks, highlighting how linguistic cues and observational experience jointly shape decision policies—a perspective that complements the more instruction-centric approaches seen in adjacent branches while addressing open questions about scalability and generalization across social contexts.

Claimed Contributions

Computational framework for social learning via joint inference over structured world models

The authors propose a Bayesian framework that treats linguistic guidance and direct experience as complementary evidence sources for inferring executable, program-like world models. This enables agents to integrate both modes of knowledge acquisition during learning in complex tasks.

10 retrieved papers
Language models as probabilistic speaker models for advice interpretation and generation

The framework leverages language models to approximate the probability that a human with specific beliefs would produce particular advice. This enables the model to both evaluate received advice under different world hypotheses and generate human-interpretable guidance for others.

4 retrieved papers
Can Refute
Language-accelerated Bayesian inference through targeted proposal distributions

The authors introduce a mechanism where language models bias the proposal of game rules during inference by converting linguistic advice into probability distributions over candidate rules. This accelerates convergence by directing inference toward theories most compatible with received guidance.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Computational framework for social learning via joint inference over structured world models

The authors propose a Bayesian framework that treats linguistic guidance and direct experience as complementary evidence sources for inferring executable, program-like world models. This enables agents to integrate both modes of knowledge acquisition during learning in complex tasks.

Contribution

Language models as probabilistic speaker models for advice interpretation and generation

The framework leverages language models to approximate the probability that a human with specific beliefs would produce particular advice. This enables the model to both evaluate received advice under different world hypotheses and generate human-interpretable guidance for others.

Contribution

Language-accelerated Bayesian inference through targeted proposal distributions

The authors introduce a mechanism where language models bias the proposal of game rules during inference by converting linguistic advice into probability distributions over candidate rules. This accelerates convergence by directing inference toward theories most compatible with received guidance.

Table of Contents