IC-079GPT-4's free-form critique generation is unreliable, containing hallucinated edges, vertex colors, and precondition states

Kaya Stechly, Karthik Valmeekam, Subbarao Kambhampati

SourceOn the self-verification limitations of large language models on reasoning and planning tasks

When asked to critique a proposed solution, GPT-4 produces feedback that is frequently wrong. In graph coloring, it hallucinates non-existent edges between same-colored vertices and misstates vertex colors, often missing the actual violated constraint. In Game of 24, it evaluates 81.6% of correct expressions to 24 but only labels 79.1% as correct, meaning it rejects valid answers even when its arithmetic is right. In planning, it hallucinates whether action preconditions are met, and in mystery blocksworld it incorrectly assumes the state of preconditions, dropping critique generation accuracy to 3% for correct plans.

Evidence
correlational
Key metric
Table A3 (Game of 24, correct expressions): correct evaluation 81.6%, correct verification 79.1%. Table A4 (graph coloring, ablated colorings with 187 errors): 24 edge hallucinations, 52 vertex hallucinations, 5 both, 24 none. Table A6 (mystery blocksworld, correct plans): binary verification 3/100 (3%), type identified 3/100 (3%), critique generation 3/100 (3%).
Caveat
Hallucination rates vary with the error rate of the input: when more edges are wrong (random colorings), the model is more likely to point to a real error. The authors note this makes the hallucination rate harder to interpret in isolation.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Failure mode
Datasets
Game of 24 (4nums.com instances 1-1000) [eval], PlanBench [eval]
Methods
VAL [validation]
Related findings
IC-078, IC-080
Extraction
automatic-extraction