---
title: 'AGI Score: Benchmarking AI Progress'
url: https://www.emergentmind.com/topics/agi-score
type: topic
---

# AGI Score: Benchmarking AI Progress

Artificial General Intelligence (AGI) Score is an aggregate metric or set of metrics designed to quantify the progress, proficiency, and safety of AI systems on their approach toward general intelligence. The concept encompasses multidomain benchmarking, cognitive profiling, safety readiness, ability generalization, and robustness, often referencing empirically validated human cognitive frameworks. The AGI Score serves both technical and policy-oriented functions: tracking advancement in AI capabilities, diagnosing cognitive deficits, evaluating safety, and guiding research, deployment, and governance.

## 1. Foundations and Definitions

Various AGI Score frameworks are grounded in human psychometrics, cognitive theory, and signal-level invariants. A widely adopted quantitative definition operationalizes AGI as the capacity of an AI system to match or exceed the cognitive versatility and proficiency of a well-educated adult, as measured using frameworks such as Cattell-Horn-Carroll (CHC) theory [2510.18212]. This approach breaks cognition down into ten domains—including knowledge, reasoning, memory, perception, and speed—each weighted equally in aggregate scoring.

The basic formula for the "CHC-style" AGI Score is:
\[ \text{AGI Score} = 10\% \cdot (K + RW + M + R + WM + MS + MR + V + A + S) \]
where each letter denotes the normalized domain score, ranging from 0–10%, yielding an interpretable percentage of human-level general intelligence.

## 2. Methodological Approaches

### Multi-domain Profiling and Psychometric Batteries
Evaluation methodologies often adapt human psychometric batteries to test individual cognitive components (e.g., Raven's Progressive Matrices for inductive reasoning, text comprehension, arithmetic) [2510.18212]. Domain scores expose "jagged" cognitive profiles in contemporary AI, with high proficiency in knowledge-intensive areas and marked deficits in foundational domains such as long-term memory storage.

### Coherence-based and Compensability-adjusted Aggregation
Traditional arithmetic mean scoring (|Compensability|) allows strengths in certain domains to mask critical failures elsewhere. A coherence-based measure [2510.20784] integrates generalized means across compensability exponents ($p$), spanning arithmetic ($p=1$), geometric ($p=0$), and harmonic ($p=-1$) regimes. The area under the curve (AUC):
\[ \text{AGI}_\mathrm{AUC} = \frac{1}{p_\mathrm{max} - p_\mathrm{min}} \int_{p_\mathrm{min}}^{p_\mathrm{max}} \text{AGI}_p\,dp \]
robustly penalizes imbalance and inter-domain dependency. For example, GPT‑5's arithmetic mean score (~58%) collapses to ~24% coherence-adjusted AUC, reflecting persistent weak domains.

### Cluster Stability and Causal Centrality Weighting
General intelligence is conceptualized as a homeostatic property cluster—a set of abilities maintained co-present under perturbation [2510.15236]. The AGI Score is refined by weighting each domain by its causal centrality:
\[ w_i^\mathrm{prior} = \lambda\,\tilde{g}_i + \mu\,\tilde{s}_i, \quad (\lambda + \mu = 1) \]
using empirical CHC loadings $\tilde{g}_i$ and mechanistic priors $\tilde{s}_i$. Cluster Stability Indices—Profile Stability (pCSI), Durable Learning (dCSI), Error-Decay (eCSI)—are computed for persistence across sessions:
\[ \mathrm{CSI} = \left(\max(\varepsilon,\, pCSI) \cdot \max(\varepsilon,\, dCSI) \cdot \max(\varepsilon,\, eCSI)\right)^{1/3} \]
Penalizing brittleness, instability, and lack of robust learning.

### Economic Impact Metrics
AGI Scores have been connected quantitatively to macroeconomic indicators. A mathematical algorithm based on Cobb–Douglas production functions relates the level of AGI technology ($AGI_n$) to real GDP ($Y_n$) [2506.03156]:
\[ Y_n = A_n \cdot K_n^a \cdot (L_n \times AGI_n)^{1-a} \]
Regression analysis reveals that a 12.5% increase in AGI level associates with a 1% increase in GDP, with a robust Pearson correlation ($r=0.854$).

## 3. Benchmarking Practices

Comprehensive AGI Score evaluation employs multidomain and multimodal approaches:
- AGIBench [2309.06495] labels each task by ability branch, knowledge category, difficulty (human-referenced accuracy), and input modality, graded over 20 categories. Multi-granularity, zero-shot testing, and auto-scoring algorithms (heuristic regex and fallback extraction) yield average, worst, best, majority-vote, and repeatability metrics aggregating into aggregate scores.
- Signal-level benchmarks such as AGITB [2504.04430] probe low-level cognitive precursors through binary sequence prediction, isolating core computational invariants (determinism, sensitivity, generalization), engineered to resist brute-force and memorization strategies. All tests must be passed for AGI-level competence.

Competitive frameworks like AGI-Elo [2505.12844] model both agent and test case difficulty as ratings in a joint probabilistic system, updating scores per "competitive interaction" and quantifying "competency gaps" in long-tail challenges.

## 4. Safety, Reliability, and Policy Dimensions

AGI Score frameworks increasingly integrate safety, control, and policy readiness:
- Composite metrics include value alignment (IRL, CIRL compliance), robustness to adversarial attacks, transparency (e.g., ReluPlex formal verification, t-SNE auditing), and corrigibility (willingness to accept shutdown or modification commands) [1805.01109].
- Public policy can directly elevate AGI Score: requirements for full disclosure, peer-reviewed safety trials, cross-checking training data (reward corruption prevention), and compliance with regulatory standards (intrinsic and extrinsic measures) enhance safety and controllability.
- Scaling laws and risk matrices account for autonomy levels in deployment (Levels 0–5), mapping performance, breadth, and autonomy to guide risk mitigation [2311.02462].

## 5. Limitations, Gaps, and Future Progress

Despite advances, current AGI Scores highlight significant limitations:
- State-of-the-art models (e.g., GPT-4: 27%, GPT-5: 58% [2510.18212]) reveal "jagged" cognitive profiles, with major deficits in foundational domains.
- Coherence-based measures expose a much lower effective score (~7% for GPT-4, ~24% for GPT-5 [2510.20784]), stressing the impact of imbalanced abilities.
- Signal-level benchmarks show that no current AI system passes all low-level cognitive tests required for AGI [2504.04430].
- High aggregate scores on benchmarks like ARC-AGI (e.g., o3 at 87.5%) are achieved predominantly via brute-force search and skill exploitation, failing the adaptability and efficiency crucial for AGI [2501.07458].
- Economic correlation models point to a lag between AGI development and realized GDP impact—posing policy and strategic challenges [2506.03156].
- Transformative AGI—defined as full substitution for human labor across nearly all valuable tasks—is judged extremely unlikely (<1% by 2043) under conservative cascading probability models [2306.02519].

## 6. Practical Applications and Directions

AGI Score frameworks provide operational metrics for:
- Diagnosing specific cognitive bottlenecks and guiding targeted research (e.g., long-term memory integration, robust reasoning, continual learning module development).
- Benchmark development—advising on multidomain, multimodal, and process-resilient testing protocols, such as persistent learning, stability under perturbation, and anti-gaming safeguards [2510.15236].
- Economic planning—quantifying investment leverage through AGI-induced productivity gains [2506.03156].
- Governance—informing global regulatory frameworks, compliance standards, and deployment readiness through transparent, reproducible scoring.

A plausible implication is that future AGI Score formulations will increasingly weigh domain coherence, stability, and real-world competence, leveraging data science methodologies (e.g., out-of-time testing, agency calibration [2510.01687]) over synthetic, easily exploited proxies.

## 7. Comparative Table of AGI Score Frameworks

| Framework / Paper           | Key Metric Type           | Domain Coverage / Weighting            |
|----------------------------|--------------------------|----------------------------------------|
| CHC-based Score [2510.18212]| Arithmetic mean of 10 domains | Equal (10% per domain), psychometric task grounding |
| Coherence-based AUC [2510.20784]| Integral of generalized means (p in [-1,1]) | Penalizes imbalance, reflects adequacy across all domains |
| Cluster Stability Index [2510.15236]| Geometric mean of persistence, learning, error-correction indices | Weighted by causal centrality, sensitivity bands |
| AGIBench [2309.06495]      | Multi-metric (average, worst, best, repeatability) | Granular branches, multimodal input, human-referenced difficulty |
| Signal-level (AGITB) [2504.04430]| All-tests-must-pass, low-level cognitive invariant compliance | Determinism, sensitivity, generalization, biological alignment |
| AGI-Elo [2505.12844]       | Competitive rating, competency gap analysis | Agent and test case difficulty, across vision/language/action |
| Economic Score [2506.03156]| Regression coefficient, Pearson's r | AGI-induced GDP increase per incremental AGI development |

The diversity of AGI Score frameworks reflects multiple priorities: cognitive breadth, coherence, economic impact, safety, and robustness. The prevailing trend is toward composite, multidomain metrics supported by rigorous benchmarking and transparent diagnostics, designed to resist superficial gaming and guide the incremental, safe advance of Artificial General Intelligence.

Source: https://www.emergentmind.com/topics/agi-score