A Simple Question
Ask an LLM: "In which country is the Eiffel Tower?" It answers "France."
Now ask two sub-questions: "In which city is the Eiffel Tower?" → "Paris." "In which country is Paris?" → "France."
The composed sub-answers match the direct answer. Normal.
But sometimes the LLM answers "France" directly, while the sub-questions yield "Paris" and "Germany"—composing to "Germany" and contradicting the direct answer. What does that mean? The model's reasoning is internally inconsistent—it isn't really "reasoning," it's making things up.
This simple check is the core idea of Operadic Consistency (OC), and its mathematical foundation is operad theory from algebraic topology.
From Intuition to Math: What Is an Operad?
An operad is a formal tool in mathematics for describing "compositional systems." Its core question: how do operations on multiple inputs compose into larger operations?
A cooking analogy:
- "Scramble eggs" is an operation with inputs {eggs, oil, salt}
- "Make a sandwich" is another operation with inputs {bread, lettuce, scrambled eggs}
- "Make a sandwich" can embed "scramble eggs"—the output of one feeds an input of the other
- A "question template" is an operation whose inputs are slots to fill
- "Decomposition" breaks a complex question into sub-questions, then recombines sub-answers
- Operadic consistency checks whether the direct answer matches the recomposed one
Operad theory precisely describes this "operations nested in operations" structure.
In LLM reasoning:
Why OC Beats Existing Methods
Existing approaches to detecting reasoning failures fall into three classes:
1. Self-Consistency: sample multiple times, check if answers agree. Problem: a model may confidently give the same wrong answer every time. 2. Semantic Entropy: measure semantic diversity of answers. Problem: only looks at output diversity, not reasoning structure. 3. P(True): ask the model "do you think you're right?". Problem: models can be confidently wrong.
OC's uniqueness: it inspects the internal structure of reasoning, not surface features of outputs. A model can look consistent (same wrong answer every time), but if its sub-reasoning fails to compose, OC catches the contradiction.
A detective analogy: self-consistency is asking a suspect the same question three times; semantic entropy counts how many versions of the story they give; P(True) asks "are you telling the truth?"; OC takes the testimony apart and checks whether the details corroborate each other.
Experimental Results: A Strong Signal Across Models and Datasets
The paper tested 17 models (12 instruction-tuned LLMs + 5 frontier reasoning models) on 5 compositional reasoning benchmarks:
OC–accuracy correlation: On 4 multi-hop QA datasets, OC rate correlates with accuracy at Pearson r ∈ [0.86, 0.94] (all p ≤ 0.0004). It is the only signal achieving r ≥ 0.85 on all 4 datasets.
For comparison, chain-of-thought self-consistency (CoT-SC) does well on HotpotQA and DROP (r = 0.93, 0.87) but collapses to r ≈ 0.45 on MuSiQue and StrategyQA.
Per-question complementarity: After controlling for CoT-SC and semantic entropy, OC still provides significant additional information (cluster-robust p ≤ 10⁻¹⁶). The result holds even against purpose-built "decomposition-aware" baselines (p ≤ 10⁻¹³).
Selective prediction: At equal compute budget (K=3), OC improves AUARC by +0.086 to +0.096 and AUROC by +0.092 to +0.164 over a tuned CoT-SC baseline, with 95% confidence intervals excluding zero.
Reasoning models: On 5 frontier reasoning models, extracting decompositions from the models' own chains of thought, equal-cost comparisons show positive gains in all 16 (dataset, budget, metric) cells, 12 with 95% confidence intervals excluding zero.
Cross-model accuracy prediction: OC rate predicts accuracy on unseen models—via leave-one-out, fitting linear regression on the other 11 models yields a mean absolute error of only 3.0 percentage points. CoT-SC (K=10, 10 inference calls) has an error of 4.6 points with a bimodal pattern.
Why This Matters
OC reveals a deep insight: LLM reasoning failures are often not random but structural.
When a model contradicts itself across sub-question reasoning, it's not an "occasional mistake"—its reasoning isn't actually composing. The direct answer and the decomposed answer travel two different paths that never communicate.
This echoes the "dual-process theory" of human cognition: System 1 (intuition) and System 2 (analysis) can give different answers. OC detects exactly that inconsistency in AI.
More importantly, OC requires no labels—no need to know the correct answer, only to check whether the model's own responses are internally coherent. This makes it nearly zero-cost in deployment: 3 inference calls (1 direct + 2 sub-questions) yield a stronger signal than 10-sample self-consistency.
An Honest Assessment
OC has so far been validated only on compositional reasoning tasks (multi-hop QA, math word problems); its applicability to open-ended generation is unclear. Not all questions have natural decomposition structures.
OC also has false positives and false negatives: a model may be OC-consistent but wrong (confidently wrong), or OC-inconsistent but right (taking a different valid reasoning path). The paper presents these cases candidly.
The operad-theoretic framework currently uses only depth-2 decompositions (one level of sub-questions); deeper tree-structured decompositions remain to be explored. But as a first step, OC already makes a convincing case: checking the internal structure of reasoning is more effective than checking its surface features.
---
Paper: Operadic consistency: a label-free signal for compositional reasoning failures in LLMs Code: https://github.com/natebottman/operadic-consistency-paper Authors: Nathaniel Bottman (Incubilate), Yinhong Liu (Cambridge), Kyle Richardson (Allen AI) Field: LLM Reasoning, Uncertainty Quantification, Mathematical Foundations