Qianfan Zhang

Jingyi Liu, Aviad Rubinstein, Ertem Nusret Tas et al.

Classical optimal auction theory assumes that bids reach the seller directly. We study how this picture changes when a revenue-maximizing intermediary controls access to the seller's auction. Motivated by blockchain auctions, online platforms, and other intermediated markets, we consider a single-item auction with independent private values and a monopolist intermediary who can decide which bidder messages are forwarded to the seller. We establish approximation guarantees and impossibility results across three timing models: seller-first, intermediary-first, and simultaneous. In the seller-first model, arbitrary deterministic seller mechanisms collapse to posted-price mechanisms, and the intermediary's best response is a shifted Myerson auction. This yields a sharp separation: for regular distributions, the seller's revenue can be arbitrarily small relative to the no-intermediary optimum, while for $α$-strongly regular distributions, posted prices recover a constant fraction of the optimum with a tight dependence on $α$. We further show that timing matters: neither Stackelberg order uniformly dominates, and simultaneous play can leave both parties unboundedly worse off than in either sequential model.

Qianfan Zhang, Tianyu Guo, Xuandi Ren et al.

We study how to scale reasoning token budgets for competitive programming through two complementary approaches: training-time reinforcement learning (RL) and test-time parallel thinking. During RL training, we observe an approximately log-linear relationship between validation accuracy and the average number of generated reasoning tokens over successive checkpoints, and show two ways to shift this training trajectory: verification RL warmup raises the starting point, while randomized clipping produces a steeper trend in the observed regime. As scaling single-generation reasoning during RL quickly becomes expensive under full attention, we introduce a multi-round parallel thinking pipeline that distributes the token budget across threads and rounds of generation, verification, and refinement. We train the model end-to-end on this pipeline to match the training objective to the test-time structure. Starting from Seed-OSS-36B, the full system with 16 threads and 16 rounds per thread matches the underlying RL model's oracle pass@16 at pass@1 using 7.6 million tokens per problem on average, and surpasses GPT-5-high on 456 hard competitive programming problems from AetherCode.