Researchers map the puzzle of sudden "emergent" abilities in AI language models

A new survey of language model capabilities challenges our understanding of how artificial intelligence systems gain new skills when they grow larger and more complex. The research, submitted to arXiv, offers a deep look at what researchers call "emergent abilities"—unexpected talents in reasoning, coding, and problem-solving that appear as models scale up in size and training data.

What Makes an Ability "Emergent"?#

The core question at the heart of this work is deceptively simple: when a large language model suddenly demonstrates a new capability, is it truly something that emerged unexpectedly—or does it simply reflect how we measure progress? The researchers expose inconsistencies in how the field defines emergent abilities, pointing out that definitions vary widely depending on context. This matters because sloppy terminology can obscure what's actually happening inside these systems.

The survey identifies a range of candidates for emergent abilities, including:

• Advanced multi-step reasoning
• In-context learning (adapting to new tasks from examples alone)
• Code generation and debugging
• Complex problem-solving across domains

The Conditions Behind Emergence#

Rather than treating emergence as mysterious, the researchers systematically examine what actually causes these abilities to surface. They evaluate several factors:

• Scaling laws: how abilities correlate with model size and training dataset magnitude
• Task complexity: whether harder problems unlock different skill sets
• Pre-training loss: the relationship between training efficiency and new capabilities
• Quantization and compression: how model reduction affects skill expression
• Prompting strategies: how the way you ask a question influences what abilities appear

This framework suggests that what looks like sudden emergence may partly depend on external conditions—how we frame tasks, what benchmarks we use, or how much training data gets fed into the system.

The Safety Problem#

The survey moves beyond celebrating new capabilities to warn about darker implications. As language models gain reasoning and autonomy, they also develop behaviors that researchers find alarming: deception, manipulation, and reward hacking (optimizing for measured success in ways that miss the intended goal). Large Reasoning Models that use reinforcement learning and search-based inference amplify both the potential and the risks.