Study: Llama 3 8B matches full fine-tuning by deleting under 0.05% of weights

This research result reframes what reinforcement learning actually changes inside a model. The team trained Llama 3 8B Instruct with RL on the GSM8K math benchmark. Instead of updating all parameters, they found that removing a tiny subset of weights, fewer than 0.05%, matched or slightly beat full fine-tuning: 83.55% versus 83.24%. The deltas fit into an 8.4 MB adapter compared with a 30.6 GB full checkpoint, roughly 3,600x smaller.

The method is a sparse intervention. Rather than learning additive low-rank updates as LoRA does, the approach identifies weights whose deletion unlocks the desired behavior. That supports a growing body of evidence that RL mostly reshapes and elicits capabilities already present from pretraining rather than teaching new ones. It also suggests customization could be stored, shipped and swapped at negligible cost.

Sparse fine-tuning competes with LoRA, QLoRA and newer delta-compression techniques. If the result holds beyond GSM8K, it could cut the cost of RL customization for enterprises and for platforms serving many per-customer variants. Researchers are excited about exactly that, according to community discussion. The Llama context is slightly ironic: Meta's own focus has shifted to the proprietary Muse family, and DigitalOcean is dropping Llama 4 Maverick from serverless inference.

The caveats are significant. GSM8K is a narrow, saturated benchmark, and an 8B model is small. Whether deletion-based updates scale to larger models, broader tasks, or capability additions rather than elicitation is untested. Watch for replication on harder reasoning suites and adoption in open fine-tuning libraries.