Ablating 80% to 90% of neurons in the Qwen 3.6-35B A3B language model improved performance on ARC-AGI benchmark puzzles

Overview Debates Related claims Topics

On the Artificial Analysis benchmark suite, Fable 5.2 cost $26.54 per task when operating on max effort settings

On the Artificial Analysis Intelligence Index, Fable 5.2 achieved a score of 65, placing 5 points ahead of Gemini Carbon, 6 points ahead of Opus 5.5, and 8 points ahead of GPT-6 Astra

On the RoboDojo Simon robotics benchmark, Fable 5.2 scored 64.9%, exceeding Opus 5.5's score by 25 points

The Neuron Scientist agent generates paragraph-length hypotheses describing the specific conditions under which an individual neuron activates

On the Remote Labor Index contractor benchmark, Fable 5.2 achieved a score of 41.3%, compared to Opus 5.5's score of 21%

On protein design benchmarks, Mythos 5.2 scored 85.1%, outperforming Mythos 5.1 by 6 percentage points

On the Artificial Analysis benchmark suite, Fable 5.2 cost $9.08 per task on high effort settings

On the TerminalBench Science 0.1 benchmark for agentic scientific research, Fable 5.2 scored 6 points ahead of GPT-6 Astra

Nick Cammarata authored the essay Neural Networks are Already Clean

Noam Shazeer proposed the SwiGLU gated multi-layer perceptron architecture in a 2020 paper

On the Artificial Analysis benchmark suite, Fable 5.2 cost $14.83 per task on extra high effort settings

The Neuron Scientist agent utilizes experimental tools including generating counterfactual sentence pairs and steering activations to refine neuron hypotheses

On the Artificial Analysis Intelligence Index, Fable 5.2 scored 62 on both the Extra High and High effort settings

Specific reasoning neurons in language models fire to modulate confidence, induce delays, trigger self-checking, or pivot problem-solving directions

On the AA Briefcase benchmark, Fable 5.2 recorded a lower Elo score than Opus 5.5

Kurt Tigges and Nick Cammarata co-authored the research paper Neurons as an Interpretable Basis

Ablating approximately 90% of individual neurons by retaining only extreme activations causes minimal performance degradation on standard language model benchmarks

On the AA Briefcase agentic knowledge work benchmark, Fable 5.2 scored approximately 5 points higher than Fable 5.1, Opus 5.5, and Gemini 4 Carbon

On the FrontierCode coding benchmark, Fable 5.2 scored 4 points ahead of GPT-6 Astra and 3.5 points ahead of Opus 5.5

Individual neurons in language model multi-layer perceptrons can express a compact set of distinct meanings rather than hundreds of unrelated meanings