IC-1328GPT-4's performance on MUSR depends on the type of neurosymbolic scaffolding: program-aided decomposition helps on structured optimization but symbolic belief tracking fails on natural language theory-of-mind

Zayne Rea Sprague, Xi Ye, Kaj Bostrom, Swarat Chaudhuri, Greg Durrett

SourceMuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning

The paper tests three neurosymbolic approaches, each using GPT-4 (or GPT-3.5) as the underlying model, on the domain that best matches their design. PAL (program-aided language models) outperforms GPT-4 CoT+ on team allocation (77.2/87.2 vs 68.4), showing that GPT-4 benefits from being asked to write and execute a program for structured optimization. In contrast, SymbolicTOM (a symbolic belief tracker) severely underperforms on object placements (23.8 vs 60.9), because its entity extraction assumes templated single-sentence facts rather than distributed natural language. Decomposed prompting is slightly worse than end-to-end CoT+ on murder mystery (77.6 vs 80.4).

Evidence
correlational
Key metric
PAL: 77.2 (zero-shot), 87.2 (1-shot) vs GPT-4 CoT+ 68.4 (team allocation); SymbolicTOM: 23.8 vs GPT-4 CoT+ 60.9 (object placements); Decomposed prompting: 77.6 vs GPT-4 CoT+ 80.4 (murder mystery)
Caveat
Each neurosymbolic method is tested on only one domain (the one it was designed for), so cross-domain generalization is not assessed. SymbolicTOM uses GPT-3.5 for state extraction, not GPT-4.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report, GPT-3.5 / ChatGPT-3.5 GPT-3.5-turbo
Methods
PAL (Program-Aided Language Models) / PaL [primary], SymbolicTOM [primary], Decomposed Prompting [primary]
Related work
SymbolicTOM [builds-on], Decomposed Prompting [builds-on]
Related findings
IC-1327
Extraction
automatic-extraction