Emergent Abilities: Real or Mirage?

Debating whether LLM capabilities appear suddenly at scale or are artifacts of non-linear evaluation metrics.

🔴 advanced5 min readllmtheory

Emergent Abilities (Wei et al., 2022) refer to capabilities (e.g. multi-step arithmetic, 3-digit multiplication, code synthesis) that are absent in small LLMs but appear sharply once model scale crosses specific parameter thresholds (e.g. N > 10^10). The 'Mirage Hypothesis' (Schaeffer et al., 2023 - NeurIPS Best Paper) argued that emergence is an illusion created by non-linear step-function evaluation metrics (e.g. Exact Match 0/1 accuracy). When evaluated with smooth continuous metrics (Edit Distance, Brier Score), capabilities scale linearly and predictably with compute.

What is an Emergent Ability?

Wei et al. (2022) observed that for certain tasks (3-digit arithmetic, IPA translation, chain-of-thought reasoning), models below 10B parameters perform at random chance level (~0%).

Once model scale crosses a critical threshold ($\sim 100\text{B}$ parameters), performance jumps sharply:

  Exact Match Accuracy (%)
  100% ┤                                                  / Sharp Emergence!
       │                                                 /
   50% ┤                                                /
       │                                               /
    0% ┴──────────────────────────────────────────────/──────────────────► Model Parameters N
       1B             10B            100B          1000B

The Mirage Hypothesis (Schaeffer et al., 2023)

Is this sharp jump a fundamental physical phase transition in neural network representations, or a measurement artifact of non-linear evaluation metrics?

  METRIC EFFECT ON SCALING CURVE SHAPE:
  - Exact Match Metric (Non-Linear 0/1):        Y = P(Correct)^k       ──► Sharp Discontinuous Jump!
  - Token Edit Distance (Continuous Linear):    Y = 1 - TokenErrorRate ──► Smooth Predictable Scaling!

The Math Behind the Mirage

Suppose predicting a 5-digit number requires 5 independent correct tokens with per-token accuracy $p$:

$$\text{Exact Match Score} = p^5$$

  Per-Token Accuracy (p)      Exact Match Score (p^5)      Perception
  ─────────────────────────────────────────────────────────────────────
  p = 0.10                    0.00001 (0.001%)             "Model cannot do task"
  p = 0.30                    0.00243 (0.24%)              "Model cannot do task"
  p = 0.50                    0.03125 (3.1%)               "Model cannot do task"
  p = 0.80                    0.32768 (32.8%)              "SHARP EMERGENCE!"
  p = 0.90                    0.59049 (59.0%)              "SHARP EMERGENCE!"

Notice that per-token capability $p$ grew smoothly and linearly from 0.10 to 0.90! The sharp jump in $p^5$ is an artifact of the exponential non-linear metric.

Practical Implications for ML Engineers

Predictable Scaling: Model capabilities scale smoothly with compute under fine-grained continuous metrics (Cross-Entropy, Perplexity, Token Distance).
Evaluations Matter: When benchmarking model performance, use continuous loss metrics alongside binary pass/fail metrics to detect gradual capability improvements in smaller models.

Say this out loud

"Emergent abilities describe capabilities that appear abruptly at scale under 0/1 exact-match metrics. The Mirage Hypothesis (Schaeffer et al., 2023) showed that emergence is often an artifact of non-linear evaluation metrics—like requiring k tokens to be 100% correct simultaneously (P^k)—whereas token-level probabilities scale smoothly and predictably with compute."

Follow-ups to expect

Are there genuine non-metric emergent abilities? Yes. In multi-step reasoning and algorithmic tasks with discrete state transitions (e.g., modular arithmetic, grokking), internal representations show non-linear topological reorganization during training.
What is Grokking in Neural Networks? The phenomenon (Power et al., 2022) where a network overfits training data with 100% train / 0% val accuracy for thousands of epochs, before suddenly generalization jumps to 100% validation accuracy long after training loss has saturated.

Check yourself

Question 1 of 3

What is an 'Emergent Ability' in Large Language Models as originally defined by Wei et al. (2022)?

More in LLMs & GenAI

See all →

Pretraining → SFT → RLHF5 min Fine-Tune vs RAG vs Prompt: Choosing5 min Retrieval-Augmented Generation5 min