Research map / Large language models / AI and society
Measuring Error Alignment for Decision-Making Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes two behavioural alignment metrics, misclassification agreement (MA) and class-level error similarity (CLES), that compare how two classifiers make errors, and correlates them with existing behavioural and representational alignment metrics across synthetic and naturalistic vision and video datasets.
Problem
Behavioural alignment metrics such as error consistency (EC) only capture whether two systems err on the same instances, ignoring which wrong classes are predicted, and they need instance-level data that may be unavailable; it is also unclear how well behavioural alignment reflects representational
Approach
MA applies multiclass Cohen's kappa to the error-agreement matrix of jointly misclassified instances, so it scores which wrong classes are chosen. CLES compares per-class error confusion distributions via a weighted Jensen-Shannon divergence, needing only class-level error counts rather than instance-level pairing.
Key results
- MA and CLES correlate with CKA in most domains; e.g. CLES vs CKA Spearman r=0.87 on ImageNet-A and MA vs CKA r=0.89 on MPII-Cooking.
- EC and MA are weakly negatively correlated globally (r=-0.23 synthetic, r=-0.29 naturalistic) but typically strongly positively correlated within domain.
- Human-human MA values are significantly lower than most model-model MA values, while EC does not show this pattern.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MA as a behavioural-alignment measure alongside CKA | Classification-system pairs · Synthetic modelvshuman subsets | Spearman's r: 0.0847 to 0.8742 | Mixed |
| Use MA as a behavioural-alignment measure alongside CKA | Classification-system pairs · Naturalistic datasets | Spearman's r: 0.7334; 0.8853; 0.802 | Helped |
| Compare behavioural alignment across domains rather than relying on a single pooled correlation | EC and MA · Synthetic and naturalistic datasets | Spearman's r: r=-0.23; r=-0.29 | Mixed |
| Use CLES to compare class-level error distributions instead of instance-by-instance joint errors | Classification-system pairs · Across domains | Spearman's r between MA and CLES: 0.20 to 0.90 | Mixed |
| Compare behavioural-alignment scores with CKA representations | Classification-system pairs · Across datasets | Spearman's r: 0.0768 to 0.8742 | Mixed |
| Use MA rather than EC alone to assess which incorrect classes systems predict | Classification-system pairs · modelvshuman | Error-pattern alignment | Helped |
| Interpret MA scores cautiously in very high-accuracy settings | Classification-system pairs · colour | Spearman's r between CKA and MA: 0.0847 | Hurt |
Surprising findings
- EC and MA are weakly negatively correlated globally, despite being positively correlated within most domains.
- CLES correlates very poorly with CKA on Epic-Kitchen (Spearman r=0.08), attributed to extreme class bias.
Limitations
- Human-human and human-machine comparisons are limited, so it is unclear whether BA metrics can proxy RA metrics for humans.
- MA may be unreliable in near-perfect-accuracy domains; CLES is affected by class bias.
- Analysis of RA vs BA correlations uses only machine-machine pairs in the main results.
Experimental setup
Datasets: modelvshuman (14 subsets) ImageNet-A MPII-Cooking Epic-Kitchen
Benchmarks: modelvshuman ImageNet-A MPII-Cooking Epic-Kitchen
Metrics: Spearman's r error consistency misclassification agreement class-level error similarity CKA SOC SOCE
Baselines: error consistency (EC) CKA
Builds on: error consistency (Geirhos et al.) Centered Kernel Alignment Cohen's kappa
Closest papers by meaning
- Position: Towards Bidirectional Human-AI Alignment (NeurIPS 2025 · similarity 0.89)
- Position: Challenges and Future Directions of Data-Centric AI Alignment (ICML 2025 · similarity 0.88)
- Quantifying Uncertainty in Error Consistency: Towards Reliable Behavioral Comparison of Classifiers (NeurIPS 2025 · similarity 0.88)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.88)
- Quantifying Misalignment Between Agents: Towards a Sociotechnical Understanding of Alignment (AAAI 2025 · similarity 0.88)
- An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.