- Rating
- 1
- Confidence
- 1
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper addresses joint ranking of ads and organic content in a feed when an item's engagement depends on its surrounding content (permutation externalities), while retaining exact dominant-strategy incentive compatibility (DSIC) for advertisers. MIRT defines a maximal-in-range (MIR) mechanism: a permutation-equivariant encoder–decoder transformer with pointer output generates a bid-independent range of K candidate feeds, and at serving time the welfare-maximising feed in the range is selected under the reported bids with VCG transfers over the same range, which yields strategyproofness by construction. The transformer is trained with on-policy policy-gradient RL (POMO-style REINFORCE with a rollout baseline) whose reward is the expected best-of-K welfare over sampled bid profiles, so the range is optimised collectively rather than feed by feed. The paper additionally bounds the pseudo-dimension of the MIRT welfare class under hard attention, giving sample complexity polynomial in transformer size and logarithmic in the range width, and evaluates the mechanism in synthetic feed environments against random/greedy MIR baselines, the non-strategyproof neural mechanisms CGA and NGA, and bid-aware beam search.
Critical strengths
- Using a bid-independent MIR range with VCG transfers is the correct route to exact DSIC for whole-feed joint ranking, where approximate-regret methods violate by about 10% of advertiser utility.
- The architecture follows the mechanism's needs (permutation-equivariant encoder, once-encoded shared pool, masking that enforces composition by construction, per-rollout conditioning for range diversity).
- Training on expected best-of-K welfare is the right objective for a range rather than a single feed, and the pseudo-dimension bound makes the range-size dependence only logarithmic.
Critical weaknesses
- The theorem certifies a hard-attention, deterministic beam-search class, whereas the deployed model samples ranges with soft attention at a tuned temperature, and nothing is proven about REINFORCE reaching the near-optimal member.
- Revenue is unanalysed and appears to collapse (every mechanism in Fig. 2 has negative revenue lift over Random+MIR), which is fatal for a mechanism intended for deployment.
- The CGA/NGA comparison is not like-for-like (re-implementations with a tuned regret penalty and the authors' own regret estimator), and bid-aware beam search beats MIRT on total welfare in about half the settings.
- The default externality γ_j = 1 + 0.5·sim(ad, previous slot) with η_below = 0 is not a whole-feed permutation externality, and all evidence comes from a bespoke synthetic simulator.
- Deployment claims are unsubstantiated — "flat latency" is asserted without hardware, memory or throughput numbers, each bid prior requires separate retraining, and there is no limitations section, code or simulator release.
Questions for authors
- The theorem certifies a hard-attention, deterministic beam-search class, whereas the deployed model samples ranges with soft attention at a tuned temperature, and nothing is proven about REINFORCE reaching the near-optimal member.
- Revenue is unanalysed and appears to collapse (every mechanism in Fig. 2 has negative revenue lift over Random+MIR), which is fatal for a mechanism intended for deployment.
- The CGA/NGA comparison is not like-for-like (re-implementations with a tuned regret penalty and the authors' own regret estimator), and bid-aware beam search beats MIRT on total welfare in about half the settings.
- The default externality γ_j = 1 + 0.5·sim(ad, previous slot) with η_below = 0 is not a whole-feed permutation externality, and all evidence comes from a bespoke synthetic simulator.
- Deployment claims are unsubstantiated — "flat latency" is asserted without hardware, memory or throughput numbers, each bid prior requires separate retraining, and there is no limitations section, code or simulator release.
MIRT: Transformers for Truthful Generative Auctions – Author Responsev
Q1 – We acknowledge there is a gap between the assumptions of the theorem and the practical implementation. Simplifying assumptions are often necessary when attempting to prove theoretical results about deep learning methods. Please note that most papers that provide theoretical results on the generalization and sample complexity of transformers make similar assumptions, as piecewise linearity is the workhorse of theoretical ML. For example, Bartlett et al. (2019) establish VC-dimension and pseudodimension bounds for networks with piecewise-linear activations, while Yang et al. (2026) establish transformer sample-complexity bounds using ReLU feed-forward layers and hard attention in place of softmax attention. Our analysis follows this tradition by providing guarantees for a precisely specified, idealised architecture.
We present additional experimental results using beam search (albeit with soft attention) in Appendix H, Table 5, which is closer to the theory, and performs worse than our ancestral decoding.
Q2 – Welfare is our primary objective, but we agree that the trade-off with revenue should be quantified, and we now do so in Appendix G.
Maximizing welfare and maximizing revenue necessarily trade off against each other to some extent (e.g., in the classic single-item auction case, the revenue-optimal auction requires a reserve price which reduces welfare).
Nonetheless, real-world ad platforms target welfare as their auction design objective and use other levers, such as reserve prices, to recover revenue. Our welfare-oriented objective is motivated by established practice and research in online advertising. Google researchers Deng et al. (2024) describe aggregate advertiser welfare as a common objective for improving channel health; Facebook researchers study welfare-maximizing allocation for content-feed advertising, with reserve prices as a separate lever for controlling revenue (Colini-Baldeschi et al., 2020, The Ad Types Problem), and motivate auction design through advertiser retention and service quality (Colini-Baldeschi et al., 2020, Envy, Regret, and Social Welfare Loss).
However, if revenue is the desired mechanism design objective, we also present several solutions. We give a virtual-welfare version of our mechanism that maximizes expected revenue over the range, and training on a welfare/virtual-welfare mix raises revenue substantially at a small welfare cost, beating NGA-Strong on both while staying DSIC (Appendix G, Table 4, Figure 6).
Q3 – We thank the reviewer for their feedback which has helped us improve our implementation of the CGA and NGA baselines. We retrained CGA and NGA with the simulator as their evaluator, MIRT’s network and three times the training budget (Appendix F.4, Table 5–Strong CGA/NGA); their welfare stayed within 0.2% of the original re-implementations while their transfer rose and regret grew from 17% to 24%, so the welfare gap to MIRT is not an artefact of our re-implementation or of the regret penalty. Additionally, ρ is tuned for CGA/NGA to keep regret low as long as it is not costing significant revenue.
Despite extensive tuning, baselines still show significant regret. RegretNet-style methods are known to be very sensitive to hyperparameters (“Auction learning as a two-player game”, Rahme et al. 2020).
We ablated a non-DSIC bid-aware version of MIRT that takes bids as input in the context (Appendix I, Table 11). This variant achieves 2.5% more welfare and a regret of 25%.
Q4 – We do provide experiments with other permutation externalities (Appendix B, Table 1). However, with our default permutation externality, we still face the central challenge in this setting, namely that one cannot simply run a position auction to maximize welfare as is possible in the classic model with separable, independent quality and position effects. We therefore think it is a reasonable default choice.
Real-world auction data is valuable and closely held by large advertising platforms, so there are no useful public datasets available. Prior papers that test on real-world data (e.g. CGA, NGA, GenAuction) do not publish engagement data. We see our simulator as a reasonable way to study the mechanism design problem, which is challenging and interesting even on synthetic data. Our methods can be applied to real-world data if and when it becomes available.
Q5.
We have now measured latency (Appendix I, Table 9): For a single example, MIRT takes 21 to 23 ms per context for every K from 8 to 512 and 29 ms at K=1024 (decoded in parallel; Single H200). Greedy+MIR grows from 0.3 to 40 ms. Training on a bid prior perturbed by 30% noise changes welfare by under 0.1% (Table 13). DSIC is preserved by construction independently of the prior. We will add a limitations section and will release code, simulator and checkpoints.