Selective Response Strategies for GenAI
Authors: Boaz Taitler, Omer Ben-Porat
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We set our model with the following parameters: T = 50, ws = 0.5, a(D) = 1 - e^(-4D) and r(p) = p^alpha for alpha R. We show our results for varying values of beta, gamma and alpha. Ideally, we would compute the optimal selective response strategy for each set of model parameters, but doing so requires using a fine discretization of the range [0, 1] for each selection of xt. To ease our computation for large values of T, we choose the optimal selective response strategy in the set of cutoff strategies. We denote by Xc the set of cutoff strategies, namely, the set of all strategies in which Gen AI remains does not answer until a specific round and fully responds thereafter. Formally, for every strategy x Xc, there exists tau T such that xt = 0 for all t < tau and xt = 1 for all t tau. For each set of model parameters, we calculate the optimal cutoff strategy to maximize Gen AI s revenue and plot the differences in Gen AI revenue and the users social welfare relative to the revenue and welfare induced by the full response strategy. Experiment Setup We report the induced revenues and welfare from 60 instances. We used a standard PC with intel Core i7-9700k CPU and 16GB RAM for running the simulations. The entire execution took roughly 1 hour. |
| Researcher Affiliation | Academia | Boaz Taitler 1 Omer Ben-Porat 1 1Technion Israel Institute of Technology, Israel. Correspondence to: Boaz Taitler <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Approximately optimal Selective Response (ASR) Input: T, A, ε Output: x 1: V (t, d) <- 0, π(t, d) <- 0 for every t ∈ [T + 1] and d ∈ {0, ε, . . . , T} 2: for t = T . . . 1 do 3: for d ∈ {0, ε, . . . , t − 1} do 4: U(y) <- 0 for every y ∈ A 5: for y ∈ A do 6: p <- y e^(βa(d)y) / (e^(βa(d)y) + e^(βws)) 7: d' <- d + (1 − p) ε 8: vd(y) <- r(p) + γV (t + 1, d') 9: end for 10: V (t, d) <- max_y vd(y) 11: π(t, d) <- arg max_y vd(y) 12: end for 13: end for 14: extract x from π starting at t = 1, d = 0 15: Return x |
| Open Source Code | No | The paper does not provide any statement or link regarding the availability of source code for the described methodology. |
| Open Datasets | No | We set our model with the following parameters: T = 50, ws = 0.5, a(D) = 1 - e^(-4D) and r(p) = p^alpha for alpha R. We show our results for varying values of beta, gamma and alpha. |
| Dataset Splits | No | The paper describes a simulation study based on a defined model and parameters, rather than using external datasets that would require explicit training/test/validation splits. |
| Hardware Specification | Yes | We used a standard PC with intel Core i7-9700k CPU and 16GB RAM for running the simulations. The entire execution took roughly 1 hour. |
| Software Dependencies | No | The paper mentions running simulations but does not specify any software dependencies with version numbers. |
| Experiment Setup | Yes | We set our model with the following parameters: T = 50, ws = 0.5, a(D) = 1 - e^(-4D) and r(p) = p^alpha for alpha R. We show our results for varying values of beta, gamma and alpha. For each set of model parameters, we calculate the optimal cutoff strategy to maximize Gen AI s revenue and plot the differences in Gen AI revenue and the users social welfare relative to the revenue and welfare induced by the full response strategy. |