Neuron Auctions: researchers propose bidding on LLM internal activations for ads

Researchers have proposed Neuron Auctions, a mechanism that lets advertisers bid on internal model activations rather than surface-level text slots when placing ads in LLM-generated conversations. The preprint, posted May 12 on arXiv, argues that existing approaches—prompt injection or fixed position slots—break semantic coherence and lack the fine-grained control needed for rigorous auction design.

The authors identify brand-specific neurons in feed-forward network layers and show that competing brands activate in approximately orthogonal subspaces. That near-independence lets them define continuous intervention budgets—neuron counts and amplification factors—as auctionable commodities. The resulting menu-based auction is strategy-proof by construction and optimizes platform revenue while penalizing interventions that degrade user experience.

What stands out

1. 01Targets internal representations, not text slots. Instead of bidding on "position 3 in the output," advertisers bid on how many brand-specific neurons to amplify and by how much. The model's natural generation process weaves the brand mention into the response.
2. 02Exploits orthogonal brand subspaces. Mechanistic interpretability reveals that different brands activate distinct, nearly orthogonal neuron clusters in FFN layers. This independence allows simultaneous, disentangled interventions without cross-brand interference.
3. 03Guarantees strategy-proofness. The continuous menu auction ensures truthful bidding—advertisers have no incentive to misreport their valuations—a property that rigid slot auctions struggle to achieve.
4. 04Prices in user experience. The platform's objective function includes an explicit user-utility penalty, dynamically pricing out overly aggressive neuron amplifications that would harm conversation quality.