Emergent Capabilities in Large Models

Emergent capabilities are qualitative behaviors that appear in large language models at or above certain scale thresholds — parameter counts, training compute, or data volume — but are effectively absent in smaller models, making them discontinuous with smooth extrapolation from smaller-scale performance curves. They matter because they challenge the standard assumption that capability scales smoothly and predictably, forcing practitioners to rethink benchmarking, safety evaluation, and deployment planning. The counterintuitive fact: a model can score near-zero on a benchmark for millions of training steps, then jump to near-perfect performance within a narrow compute window — a phase transition more reminiscent of physics than engineering.

💡 The Core Insight

The central insight is that language model training implicitly learns a combinatorial hierarchy of sub-skills, and a capability only manifests when enough constituent sub-skills are simultaneously present — meaning the capability's "unlocking" looks sudden from the outside even though each sub-skill was learned gradually. Think of it like assembling a lock: individual pins move continuously, but the lock only opens when all pins align at once. This means you cannot reliably predict the presence of a capability by interpolating from smaller models, and you cannot assume a capability is absent just because it wasn't measured in a smaller checkpoint.

⚠️ Common Misconception

The most common mistake is conflating emergence with unpredictability of scale, concluding that "we simply cannot know what large models will do." The correct mental model is more precise: emergence is a statement about metric discontinuity under a given evaluation protocol, not about fundamental unpredictability of the underlying loss. As Schaeffer et al. (2023) showed, many apparently emergent capabilities disappear and become smooth curves when you switch from a discontinuous metric (exact-match accuracy) to a continuous one (log-probability of the correct token). What looks like emergence is sometimes an artifact of how we measure, not an intrinsic property of the model — practitioners must always ask whether their benchmark metric is linear in the latent capability being measured.

How It Works

The Phase Transition Framing

Let $C$ be compute (FLOPs), $N$ be parameter count, and $D$ be data tokens. The Neural Scaling Laws literature (Hoffmann et al., Chinchilla) establishes that cross-entropy loss $\mathcal{L}$ decreases smoothly:

\quad \mathcal{L}(N, D) \approx \mathcal{L}_\infty + \frac{A}{N^\alpha} + \frac{B}{D^\beta}

Explanation: The test loss decomposes into an irreducible floor plus two power-law terms that shrink independently as parameters and data grow. Here $\mathcal{L}_\infty$ is the entropy of the data distribution, $A, B$ are dataset/architecture constants, and $\alpha, \beta \approx 0.07$ are empirically measured exponents. The loss is smooth — but the downstream task accuracy is not, because accuracy is a nonlinear function of the log-probability gap between correct and incorrect completions.

From Smooth Loss to Sharp Transitions

Consider a $k$ -shot classification task where the model must assign higher probability to the correct completion $y^*$ than to $K$ distractors. Define the accuracy indicator:

\quad \text{Acc} = \mathbf{1}\!\left[\log p_\theta(y^* \mid x) > \max_{j \neq *} \log p_\theta(y_j \mid x)\right]

Explanation: Accuracy is a hard thresholding function — it jumps from 0 to 1 the moment the correct answer's log-prob crosses the maximum distractor's log-prob. Because $\mathbf{1}[\cdot]$ is a step function, even a tiny continuous improvement in $\log p_\theta$ can cause accuracy to jump from chance to perfect when the model crosses the decision boundary. Under a continuous scoring metric $s = \log p_\theta(y^*) - \max_j \log p_\theta(y_j)$ , no phase transition is visible.

A Multi-Skill Composition Model

Suppose capability $T$ requires $m$ sub-skills $\{s_1, \ldots, s_m\}$ , each learned with probability $p_i(C)$ that increases smoothly with compute. If skills are approximately independent:

\quad P(\text{capability } T \text{ present} \mid C) = \prod_{i=1}^{m} p_i(C)

Explanation: The joint probability of all sub-skills co-occurring is the product of individual probabilities, so it rises as a steep sigmoidal curve even if each factor rises gradually. As $m$ grows large, the product $\prod_i p_i$ transitions from near-zero to near-one over an increasingly narrow compute window — giving the appearance of a sharp phase transition. This is the combinatorial emergence hypothesis proposed by Arora & Goyal (2023).

Information-Theoretic Bound on Emergence

Let $H(T)$ be the task entropy (bits required to solve the task). The minimum model capacity (in nats of mutual information with training data) needed to reliably perform $T$ obeys:

\quad I(\theta; \mathcal{D}) \geq H(T) - H(T \mid \mathcal{D})

Explanation: A model must encode at least as much information about the data as the task demands, minus the information already in the task's prior — setting a hard lower bound on required scale. Below this threshold, no amount of fine-tuning can reliably elicit the capability; above it, the capability is latent and can be unlocked by prompting or instruction tuning.

Measuring Emergence Empirically

A widely used operational definition: a capability is emergent if its accuracy $a(C)$ satisfies

\quad \left.\frac{d^2 a}{dC^2}\right|_{C=C^*} \gg 0 \quad \text{and} \quad a(C < C^*) \approx \text{random}

Explanation: Emergence is characterized by a positive second derivative (upward curvature) at some critical compute $C^*$ , with near-chance performance below that threshold. In practice, practitioners estimate this by evaluating checkpoints on a log-spaced compute grid and looking for inflection points; tools like BIG-Bench provide standardized benchmarks for this purpose.

researcher · Wei Jason | b. 1992 | Formal Definition of Emergent Capabilities

Jason Wei is a research scientist at Google Brain (later OpenAI) whose 2022 paper "Emergent Abilities of Large Language Models" (with Tay, Bommasani, and others) provided the first systematic empirical survey of emergent capabilities across 137 tasks and multiple model families. His work established the now-standard operational definition — capability absent below a scale threshold and present above it — and catalogued abilities including multi-step arithmetic, logical reasoning, and chain-of-thought without explicit training. A surprising fact: Wei completed his PhD at Dartmouth in computational biology before pivoting entirely to NLP scaling research.

researcher · Sanjeev Arora | b. 1968 | Theoretical Foundations of Emergence

Sanjeev Arora is a professor at Princeton and co-director of the Princeton Language and Intelligence initiative, known for his foundational work in computational complexity and, more recently, the theory of deep learning. His 2023 collaboration with Abulhair Saparov proposed the multi-skill composition model (equation [3] above), offering the first mathematically grounded explanation of why emergence appears sharp even when underlying learning is gradual. Arora is also famous in theoretical CS for his near-linear-time approximation algorithm for the Euclidean TSP, a result completely unrelated to AI that won him the Gödel Prize.

Figure

Emergent Capability Phase Transition

Figure

Multi-Skill Composition and Emergence

Intuition

🧠 The Mental Model

Imagine you are assembling a jigsaw puzzle in a dark room with a slowly brightening light. Each piece represents a sub-skill the model is learning. For most of the assembly process, even though you're placing pieces correctly, the picture looks like noise — no recognizable image exists yet. Then, at some threshold of brightness and puzzle completion, the image suddenly snaps into recognizable focus. You didn't change your strategy; you didn't add a special piece; the picture just crossed a threshold of coherence. This is emergence. The model has been quietly accumulating relevant representations throughout training, but the downstream capability only becomes evaluable — only "snaps into focus" — once enough pieces are simultaneously in place. The key implication for practitioners: training a model twice as long may not give you twice the capabilities — it may give you zero new capabilities, and then suddenly three new ones in the next 20% of training.

Historical Development

⏳ History

2001

Shannon's channel capacity work, though older, provides the theoretical basis for equation [4] — information must pass through the model to reach task performance.

2017

Vaswani et al. introduce the Transformer in "Attention Is All You Need," creating the architecture whose scale would eventually produce observed emergence.

2020

Brown et al. publish GPT-3 (175B parameters), the first model where qualitatively novel few-shot behaviors — arithmetic, translation, code generation without task-specific training — are systematically observed and documented.

2021

Bommasani et al. coin the term "foundation model" in the Stanford CRFM report, framing large pretrained models as substrates for emergent downstream capabilities across diverse tasks.

2022

Wei et al. publish "Emergent Abilities of Large Language Models" in TMLR, providing the first systematic, multi-model empirical survey across 137 BIG-Bench tasks and defining emergence operationally.

2022

Wei et al. (separately) publish "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models," demonstrating that multi-step reasoning — itself an emergent capability — can be elicited through prompting format changes.

2023

Schaeffer, Miranda & Koyejo publish "Are Emergent Abilities of Large Language Models a Mirage?", arguing that many apparent phase transitions vanish under continuous metrics, fundamentally reframing the debate.

2023

Arora & Goyal propose the multi-skill composition theoretical framework, providing the first mathematically grounded explanation consistent with both smooth loss and sharp accuracy transitions.

2024

Emergence research expands to multimodal models (GPT-4V, Gemini), where capabilities like optical character recognition, diagram interpretation, and spatial reasoning show analogous scale-dependent onset behaviors.

Why It Matters

🔬 Real-World Impact

GPT-4 (OpenAI, 2023)Demonstrated emergent performance on the bar exam (90th percentile), AMC math competitions, and AP-level science — tasks on which GPT-3.5 performed near chance — attributable to a scale jump of roughly one order of magnitude in effective compute, illustrating that capability planning cannot rely on interpolation from smaller checkpoints.

BIG-Bench (Google, 2022)A benchmark suite of 204 tasks explicitly designed to probe emergent capabilities, used by Google PaLM, DeepMind Chinchilla, and OpenAI GPT-4 teams to track capability onset thresholds; it revealed that ~25% of tasks show sharp emergence while the rest scale smoothly.

Chain-of-Thought PromptingCoT reasoning — the ability to generate intermediate reasoning steps that improve final answers — is itself an emergent capability, absent in models below ~60B parameters; this directly motivated the instruction-tuning and RLHF pipelines in ChatGPT and Claude that teach models to reason step-by-step.

AI Safety Red-TeamingBecause emergent capabilities can appear suddenly and without explicit training signal, organizations like Anthropic and DeepMind run continuous capability evaluations on intermediate checkpoints (e.g., evaluating for deceptive reasoning or autonomous replication) as part of their model cards and Responsible Scaling Policies — a direct response to the unpredictability of emergence.

Chinchilla Scaling Laws (Hoffmann et al., 2022)By showing that GPT-3 was significantly undertrained relative to its parameter count, the Chinchilla work allowed practitioners to reach the same capability thresholds with 4× fewer parameters (70B vs 280B) at the same compute budget — effectively lowering the scale barriers at which emergence occurs.

Code Generation (GitHub Copilot / GPT-4 Codex)The ability to generate syntactically and semantically correct multi-function programs is an emergent capability first observed in Codex (12B parameters) that was absent in GPT-2 (1.5B); this directly powers Copilot's production system serving millions of developers.

Multimodal Emergence (Gemini Ultra, GPT-4V)Spatial reasoning over novel diagrams, OCR in low-resource languages, and scientific figure interpretation all exhibit emergent onset in vision-language models above roughly 10^24 FLOPs training compute, motivating the current generation of large multimodal training runs.