PaperlayerView exact revision on arXiv

Source-linked AI summary

LangBP: Language-Guided Reasoning and Acting for Joint Bidding and Pricing

Jiaqi Ding, Chuan Yang, Linghui Meng, Shengsheng Niu, Jie He, Zhangang Lin, Ching Law, Xiaolin Fang

arXiv:2608.30343v1cs.GT

TL;DR

Auto-bidding must maximize long-horizon conversion value under budget and KPI constraints, while existing language-guided methods do not explicitly model strategy-induced state changes or balance updates across similar execution effects. LangBP addresses these gaps with S-DT target-state mediation and EGPO with CEV-based effect grouping. Across AuctionNet variants it outperforms evaluated baselines, with online deployment gains, while CEV remains observational and later targets use causal reward updates rather than logged future suffixes.

  • Problem

    Existing language-guided methods lack explicit state-change modeling and can assign imbalanced updates when different instructions produce similar execution effects.

  • Method

    LangBP uses S-DT to predict target states and recover joint actions through inverse dynamics, while EGPO uses CEV and groups policy updates by execution effect.

  • Results

    LangBP achieves higher Scores than evaluated numerical, generative, and language-guided baselines across AuctionNet and AuctionNet-Sparse, with online A/B-test gains in ad value and target achievement.

  • Takeaways & Limitations

    Explicit target-state mediation and effect-grouped feedback provide complementary evidence for LangBP’s design choices in language-guided joint bidding and pricing.

  • Takeaways & Limitations

    CEV provides observational, non-causal ranking scores, and later stages exclude logged future suffixes from policy optimization and semantic alignment.

Abstract

from arXiv · show

Auto-bidding is a long-horizon sequential decision problem for maximizing conversion value under budget and key performance indicator (KPI) constraints. Recent work extends this task from bidding alone to joint bidding and pricing, where a policy controls bidding decisions and pricing corrections. Existing methods mainly rely on numerical trajectory modeling, which offers limited support for interpreting campaign context and expressing high-level strategies. Large language models (LLMs) can complement this paradigm with their reasoning capabilities. However, existing language-guided methods have two limitations. First, they condition actions on language strategies without modeling the corresponding state changes, making it difficult to distinguish errors in strategy understanding from errors in action generation. Second, different instructions can produce similar execution effects, leading to imbalanced policy updates across effects. We propose LangBP, a hierarchical framework for language-guided joint bidding and pricing. LangBP's Semantic Decision Transformer (S-DT) predicts target states from the instruction and the trajectory history, then recovers the joint action via inverse dynamics. We further propose Execution-Grouped Policy Optimization (EGPO), which scores candidate effects with a Context--Effect Verifier (CEV) and balances policy updates across effect groups. Experiments on AuctionNet show that LangBP outperforms strong baselines, and online A/B tests further demonstrate business gains in real-world deployment on a large-scale e-commerce platform.

1 Introduction

Auto-bidding extends from numerical bidding control to joint bidding and pricing, while language-guided methods seek to express campaign strategies. LangBP addresses missing state-change modeling and effect-imbalanced updates with hierarchical execution and grouped policy optimization.

  • Motivation: Auto-bidding maximizes conversion value under budget and KPI constraints through long-horizon sequential decisions.Each decision affects future spending, constraint status, and available traffic.
  • Motivation: Joint bidding and pricing separates future traffic acquisition from correction of accumulated constraint deviations.Bids affect future traffic, while pricing corrections reduce accumulated deviations.
  • Challenges: Existing language-guided methods directly predict continuous actions without explicitly representing the state changes corresponding to language strategies.Action-level supervision does not directly distinguish strategy-understanding errors from action-generation errors.
  • Challenges: Multiple instructions can induce similar execution effects, causing those effects to receive disproportionate optimization weight within sampled policy updates.This can weaken learning signals for other valuable effects.
  • LangBP: LangBP uses S-DT to predict target states from instructions and trajectory history, then recovers joint bidding–pricing actions through inverse dynamics.The target-state representation connects language strategies to continuous control.
  • LangBP: EGPO groups candidates by execution effect, uses CEV for context-dependent feedback, and balances policy updates across effect groups.The framework combines a high-level instruction policy with a low-level continuous executor.

2 Related Work

Related work develops numerical, generative, language-guided, and state-mediated approaches to auto-bidding and execution feedback. These lines motivate combining language reasoning with intermediate state representations and balanced candidate optimization.

  • Numerical and generative methods: Numerical and generative decision models learn auto-bidding policies from logged trajectories, with Decision Transformer serving as a representative offline continuous-control backbone.DiffBid is another generative approach that predicts future states and recovers bids through inverse dynamics.
  • Joint bidding and pricing: Joint bidding–pricing methods couple bidding decisions with post-auction pricing corrections to pursue future value while repairing accumulated KPI deviations.JD-BP separates future traffic acquisition from accumulated-deviation repair.
  • Language-guided control: Language-guided methods condition bidding or execution on task descriptions, histories, semantic strategies, or high-level reasoning.Examples include SemBid, DecisionLLM, SAGE, LBM, and AIGB-R1.
  • Execution feedback: Language-policy optimization uses numerical execution feedback from offline Q-functions, simulators, trajectory evaluators, or relative feedback across sampled responses.GRPO and GiGPO extend relative-feedback optimization to multiple responses or sequential language agents.

3 Problem Formulation

The formulation treats joint bidding and pricing as a constrained sequential-control problem under uncertain future opportunities. It defines the executable action space, history-based state, reward, return targets, and language-conditioned hierarchy used by LangBP.

  • 3 Problem Formulation: The problem formulation covers joint bidding and pricing with full future-opportunity knowledge, then recasts it as language-guided sequential control when opportunities are unknown.The two settings are introduced in Sections 3.1 and 3.2.
  • 3.1 Joint Bidding and Pricing Correction: Each opportunity has an estimated value, auction payment, budget, and KPI outcomes with target bounds defining the constrained bidding problem.Under perfect future-value knowledge and a stationary market, the optimal policy has a closed-form linear bid.
  • 3.1 Joint Bidding and Pricing Correction: Bidding-only control must use the same lever to maximize future value and repair accrued constraint violations, creating a conflict between these objectives.Prediction error, conversion delay, and nonstationary competition make violations unavoidable in practice.
  • 3.1 Joint Bidding and Pricing Correction: Joint bidding–pricing introduces a history-dependent payment correction that separates repair of past deviations from pursuit of future value.Related mechanism-design results show that history-dependent bank-account corrections can dominate transaction-independent pricing in revenue and social welfare.
  • 3.1 Joint Bidding and Pricing Correction: JD-BP applies a bounded non-positive post-auction pricing correction to draw down a tracked historical deficit through the payment channel.The correction changes the auction payment c_t into the actual charge p_t = c_t + y_t.
  • 3.1 Joint Bidding and Pricing Correction: The joint action combines a continuous bidding parameter with a pricing correction, whose feasible request is projected after settlement and whose bidding adjustment is measured relative to its preceding value.This joint action is the executable interface for LangBP.
  • 3.2 Language-Guided Sequential Control: Because closed-form control assumes exact future values and a stationary market, the practical problem uses observed campaign history to choose actions and transition between states.The environment emits an interval reward and a next state after each feasible joint action.
  • 3.2 Language-Guided Sequential Control: Rewards are interval conversion value with hinge penalties for budget and payment/KPI violations, preserving the constrained campaign objective while shaping training.The state summarizes delivery history, remaining budget, and constraint status.

4 Methodology

LangBP builds a hierarchical language-guided controller for joint bidding and pricing: a teacher-derived structured instruction is executed through target-state prediction and inverse dynamics, while CEV-guided EGPO updates the instruction policy by execution effects.

  • Structured instruction generation: LangBP trains a smaller instruction policy on structured outputs containing a strategy explanation, bidding direction, and pricing strength, while the teacher is used only for offline data construction.The explanation distinguishes context-dependent strategies that share categorical controls, and deployment uses the student policy with S-DT.
  • Semantic Decision Transformer: S-DT predicts the next bidding and pricing target state from language and numerical history, then recovers the feasible joint action through inverse dynamics.The inverse model does not read the instruction directly, so language influences actions through the predicted target state.
  • Semantic alignment: Observed-transition training jointly enforces target-state prediction, inverse-action recovery, and dynamics consistency using logged next states and actions.The dynamics objective checks whether the recovered action realizes the planned target, while inverse and dynamics models are pretrained from logs and frozen before semantic alignment.
  • Semantic alignment: Counterfactual instructions provide instruction and dynamics supervision but receive no state or action target because their counterfactual next states are unobserved.Only examples whose predicted actions remain within validation-selected local support of logged behavior are retained.
  • Context–Effect Verifier: CEV ranks candidate execution effects observationally using local-support filtering, matched historical pairs, and context-conditioned scoring.It restricts effects to behavior-supported contexts and scores both effects under both matched contexts, while explicitly remaining non-causal under unobserved confounding.
  • Execution-Grouped Optimization: EGPO groups instructions with similar supported effects and assigns equal outer weight to active groups, preventing repeated candidate effects from receiving greater total optimization weight.The grouping is applied after frozen S-DT maps sampled instructions to effects and unsupported effects are filtered.

5 Experiments

Experiments evaluate LangBP across offline AuctionNet settings, controlled ablations, component diagnostics, and an online tROI deployment test. LangBP achieves the best reported benchmark performance, while its state-mediated execution and effect-grouped feedback show the intended behavior and improve platform metrics.

  • 5.2 Overall Performance: LangBP achieves the highest Score in all ten AuctionNet and AuctionNet-Sparse settings across five budget ratios.On AuctionNet, it improves over JD-BP by 4.2%–6.5%; on AuctionNet-Sparse, it exceeds the corresponding strongest baseline by approximately 2.0%–3.3%.
  • 5.2 Overall Performance: LangBP achieves higher Scores than LBM and SemBid in every evaluated setting.
  • 5.3 Ablation Study (RQ2): Post-training with execution feedback improves over supervised initialization by 8.2%–11.1% across five budget ratios.LangBP exceeds Direct Action by 4.1%–6.6%, improves over Relative-Q Feedback by 2.8%–5.3%, and gains 1.4%–3.8% over w/o Effect Grouping.
  • 5.4 Component Analysis (RQ3): S-DT reaches 95.5% macro joint instruction compliance versus 89.2% for Direct Action across the 3×3 instruction combinations.Worst-cell compliance is 93.9% for S-DT and 85.0% for Direct Action, with S-DT more compliant across all nine combinations.
  • 5.4 Component Analysis (RQ3): CEV reaches 85.7% pairwise preference accuracy, compared with 82.4% for Single-Tower CEV and 72.6% for Relative-Q.Accuracy falls to 58.2% after support-preserving context permutation; these results concern observational ranking within behavior-supported regions, not causal estimation.
  • 5.4 Component Analysis (RQ3): EGPO keeps relative update-magnitude medians between 1.00 and 1.03 with near-zero correlation to group multiplicity.Candidate-wise medians rise from 0.88 for singleton groups to 1.92 for groups with at least three candidates, with Spearman correlation 0.56; positive redundant-effect bins contain 68.5% of eligible sets.
  • 5.5 Online A/B Test: In a seven-day online tROI A/B test, LangBP improves Ad Click by 2.59%, Ad Cost by 1.86%, Ad Value by 3.34%, and Achievement by 1.83 percentage points over production.Ad Value grows more than Ad Cost, alongside higher Achievement.

6 Conclusion

LangBP is a hierarchical framework for language-guided joint bidding and pricing. Its S-DT, CEV, and EGPO components support explicit target-state execution, effect-aware feedback, and balanced policy updates, with gains reported offline and online.

  • 6 Conclusion: LangBP combines hierarchical language-guided reasoning with joint bidding and pricing.The framework includes a high-level instruction policy and low-level continuous execution.
  • 6 Conclusion: S-DT maps instructions to explicit target states and recovers joint actions through inverse dynamics.
  • 6 Conclusion: CEV provides context-dependent preferences over behavior-supported execution effects, while EGPO balances policy updates across effect groups.
  • 6 Conclusion: LangBP achieves higher Scores than evaluated numerical, generative, and language-guided baselines on AuctionNet and AuctionNet-Sparse.
  • 6 Conclusion: Online A/B testing reports gains in ad value and target achievement on a large-scale advertising platform.
  • 6 Conclusion: CEV is limited to observational evidence in behavior-supported regions.The paper identifies online feedback under distribution shift as future work.

A Additional Results

Additional results compare LangBP in a controlled action space and report the corresponding AuctionNet-Sparse evaluation table.

  • Action-space-controlled comparison: Table 4 compares full LangBP with bidding-only GUIDE, LBM, SemBid, and LangBP w/o Pricing on AuctionNet-Sparse.
  • Action-space-controlled comparison: The bidding-only comparison keeps the data split and replay evaluation unchanged while setting the pricing action to zero.

B Implementation Details

The implementation fixes validation-selected settings, specifies the data and model configuration, and details support checks, pricing execution, and CEV/EGPO parameters.

  • B Implementation Details: Validation-selected hyperparameters and checkpoints are fixed before final evaluation.
  • Data and models: Experiments use AuctionNet and AuctionNet-Sparse, with GPT-5.5 as teacher and Qwen-3.5-0.8B as the instruction policy.
  • SFT and S-DT: SFT and S-DT use specified percentile-based labels, fixed history and architecture settings, and separate pretraining and alignment learning rates.
  • CEV and EGPO: CEV uses suffix horizon 12 and five advertiser folds, while EGPO samples eight candidates and applies PPO-based optimization.
  • Pricing-action execution: Pricing corrections are constrained by a refund-rate cap and minimum charge; losing impressions receive zero price and correction.
  • Preprocessing and support checks: Counterfactual actions undergo K_a-nearest-neighbor support checks, and retained CEV pairs use context-similarity weighting.

C.1 Five-Run Results

Five-run analyses report dispersion for LangBP scores and ablations, with effect grouping producing consistently positive matched-seed differences across budget ratios.

  • C.1 Five-Run Results: Tables 6 and 7 report dispersion across the same five runs used for the means in Tables 1 and 2.
  • C.1 Five-Run Results: 0.80±0.17 to 0.90±0.19 are the matched-seed LangBP-minus-w/o-Effect-Grouping Score differences across the five budget ratios.The reported differences are 0.80±0.17, 0.90±0.22, 0.60±0.14, 0.80±0.20, and 0.90±0.19 from 50% to 150% budget.
  • C.1 Five-Run Results: The LangBP-minus-w/o-Effect-Grouping difference is positive for all five seeds at every budget ratio.

D Component-Analysis Details

This component analysis examines held-out pairwise preference accuracy, grouping-threshold sensitivity, and how candidate-wise versus EGPO weighting distributes update mass across execution-effect groups.

  • Held-out preference accuracy: All four held-out preference-accuracy comparisons use 4,231 validation pairs from 10 advertisers, with 95% confidence intervals reported for the Figure 4 order.Support filtering retains 81.6% of sampled candidates.
  • Held-out preference accuracy: Support regions are defined by training-set tertiles of the worst cross-context support distance between the two effects in each matched pair.
  • Grouping sensitivity: At thresholds 0.8𝜖𝑔, 𝜖𝑔, and 1.2𝜖𝑔, repeated-effect rates increase from 65.5% to 71.5%, while average distinct groups decrease from 5.46 to 5.05.The sensitivity check keeps eligible contexts and supported candidates fixed across thresholds.
  • Update-mass balance: Candidate-wise weighting assigns supported candidates outer weight 1/N_t, while EGPO assigns candidates in group G_t,k weight 1/(K_t|G_t,k|).Contexts with fewer than two active groups or zero total coefficient are excluded.
  • Update-mass balance: Candidate-wise update mass rises with group multiplicity, whereas EGPO remains nearly uniform across groups.Candidate-wise medians increase from 0.88 to 1.92 with Spearman correlation 0.56; EGPO medians stay between 1.00 and 1.03 with correlation −0.0023.
  • Update-mass balance: The update-mass analysis covers 5,686 active groups from 1,194 held-out contexts across 48 advertisers, including singleton and multi-candidate groups.Figure 5(b) subsamples multiplicity bins for visibility, while medians and correlations use all groups.
Loading 2608.30343v1…