Selective Neuron Amplification for Training-Free Task Enhancement
This addresses model failures due to weak activation rather than lack of capability, offering a potential enhancement for users of large language models, though it appears incremental as it builds on existing inference-time adjustments.
The paper tackled the problem of large language models failing on tasks they seem to understand by proposing Selective Neuron Amplification, a training-free method that amplifies task-relevant neurons during inference, which helps when the model is uncertain but has low effect when confident.
Large language models often fail on tasks they seem to already understand. In our experiments, this appears to be less about missing knowledge and more about certain internal circuits not being strongly activated during inference. We explore Selective Neuron Amplification, which increases the influence of task relevant neurons without changing the model's parameters. The method works at inference time and does not permanently alter the model. SNA helps mainly when the model is uncertain, while having low effect when the model is already confident. This suggests that some model failures are due to weak activation rather than lack of capability.