credit assignment

In dynamic mechanism design, traditional static contracts fail to incentivize forecasters to sustain learning effort over time. Method: This paper pioneers the integration of Bayesian learning dynamics into mechanism design, constructing a dynamic reward mechanism that depends jointly on outcomes and reports. Contribution/Results: Theoretical analysis demonstrates that information structure—particularly signal predictability—determines the optimal reporting format: restricted summary reports induce insufficient learning when signals are either non-fully revealing or unpredictable, whereas flexible reporting significantly enhances information acquisition efficiency. The paper fully characterizes the structural properties of effort-maximizing contracts, thereby transcending the static incentive paradigm. It establishes a novel theoretical foundation for dynamic information acquisition mechanisms and offers actionable insights for practical implementation.

This paper investigates intervention-driven causal representation learning (CRL) for nonparametric latent causal models, where the mapping from latent variables to observations may be linear or nonlinear. It addresses identifiability and algorithmic realizability of both latent causal variables and the underlying causal graph. Methodologically, it establishes the first theoretical connection between score functions (i.e., gradients of log-densities) and CRL; introduces general identifiability conditions that do not require intervention environment labels—significantly relaxing standard intervention assumptions; and proposes a score-matching-based optimization framework supporting both stochastic hard and soft interventions, as well as latent graph structure recovery. Theoretically, full identifiability is guaranteed with a single hard intervention in the linear case and two hard interventions in the general nonlinear setting. Extensive experiments on synthetic and image data demonstrate accurate recovery of both latent causal variables and the causal graph structure.

Existing interpretability methods for black-box image classification models suffer from insufficient causal grounding and fail to distinguish genuine causal features from spurious correlations induced by unobserved confounders. Method: This paper proposes an explainability framework grounded in interventionist counterfactual causal reasoning. It constructs a causal graph model incorporating unobserved confounders, integrates structure learning with latent variable modeling to abstract raw pixels into high-level semantic features, and identifies true “difference-makers”—features whose counterfactual interventions alter predictions—under arbitrary unmeasured confounding. Contribution/Results: This work introduces the first systematic application of counterfactual causal explanation to black-box model auditing, enabling verifiable identification of causal drivers. Experiments on image classification tasks demonstrate significant improvements in causal feature identification accuracy, thereby supporting rigorous algorithmic attribution analysis and trustworthy model evaluation.

This work addresses the conflation of causation and correlation in time-series classification interpretability, proposing the first concept-level segment-oriented causal attribution framework. Methodologically, it introduces a high-fidelity counterfactual generator based on diffusion models, enabling concept-level interventions on predefined temporal segments and rigorously estimating their causal effects on classification outcomes. Unlike conventional correlation-based attribution methods, this framework explicitly models time-series segmentation from a causal inference perspective, distinguishing causal pathways from spurious statistical associations. Experiments across multiple time-series classification tasks demonstrate that the proposed causal attribution reliably identifies critical decision-relevant segments, whereas mainstream correlation-based approaches frequently overlook true causal mechanisms—leading to systematic misattribution. The framework thus advances interpretable time-series classification by grounding explanations in causal reasoning rather than mere association.

This paper addresses the efficient and accurate pricing and sensitivity (Greeks) computation for path-dependent derivatives with early-exercise features—such as Asian, lookback, and callable warrants. We propose a novel modeling framework grounded in the signature representation of the underlying price process. Methodologically, we integrate stochastic feedforward and recurrent neural networks with signature path encoding, and innovatively employ Chebyshev polynomial interpolation to construct an efficient, differentiable Delta and Gamma computation scheme. Compared to conventional least-squares Monte Carlo (LSM) and partial differential equation (PDE) approaches, our method achieves millisecond-level pricing across all three derivative classes, with Greeks errors under 0.5%. This substantially improves the accuracy–speed trade-off. The framework is scalable to high-dimensional, non-Markovian path-dependent settings, providing a computationally tractable paradigm for real-time hedging and risk management of complex structured products.

This work addresses the mismatch between training and inference in implicit process reward models, which leads to unreliable token-level reward signals. To resolve this, the authors propose the Implicit Prefix Value Reward Model (IPVRM), which directly models a prefix-conditional value function to estimate the probability of ultimately generating a correct answer and leverages temporal difference advantage estimation to produce reliable step-level signals. Furthermore, they introduce Distributional Reinforcement Learning (DistRL), which performs dense counterfactual updates on both sampled and high-probability candidate tokens without requiring additional rollouts, enabling efficient optimization. Evaluated on ProcessBench, IPVRM significantly improves step-wise verification F1 scores, and when combined with DistRL, consistently enhances downstream reasoning performance.

Traditional probabilistic models are typically trained using task-agnostic log-loss, which often yields propensity scores with large errors, high bias, and high variance in boundary regions—particularly detrimental in causal inference tasks such as inverse probability weighting. This work proposes a general framework that, for the first time, integrates the error structure of downstream tasks into the design of strictly proper scoring rules. By aligning the local curvature of the scoring rule with that of the target loss, the authors derive a closed-form loss function tailored for average treatment effect estimation, along with its associated canonical probability mapping, enabling end-to-end task-oriented training. The approach is compatible with both neural networks and gradient boosting models and consistently outperforms standard log-likelihood and covariate balancing methods across multiple causal inference benchmarks, substantially improving estimation accuracy and stability.

This work investigates whether reinforcement learning agents truly acquire the underlying task states when achieving high rewards or merely exploit reward shortcuts. To disentangle reward acquisition from state representation, the authors construct controllable symbolic sequence tasks grounded in implicit deterministic finite automata (DFAs) with sparse terminal rewards. They introduce a white-box probing method that, for the first time, distinguishes between perceptual gaps and planning gaps, revealing that the linguistic structure of task families effectively predicts learning failure. Evaluated on 153 novel DFA-based tasks, this structural predictor achieves an accuracy of 0.86 in anticipating perceptual gaps. Experiments further demonstrate that while strong optimizers can partially recover latent states, they exhibit high variance, and critically, high reward does not necessarily imply genuine task understanding.

This work addresses the challenge of effectively transferring benchmark evaluation knowledge to novel few-shot tasks. It proposes a transfer learning and model selection aggregation method grounded in a weak monotonicity assumption: models that perform better across multiple source benchmarks tend to also perform better on the target task. The approach prunes the model class, models approximate weak monotonic relationships between tasks, and employs an adaptive hedging strategy over the performance frontier set to balance performance discrepancies among candidate models. Theoretical analysis demonstrates that this framework yields statistical gains under the weak monotonicity condition, and empirical results confirm its significant superiority over existing methods in few-shot scenarios.

This work addresses the challenge of causal inference with continuous-time marked point process data, for which existing methods lack a suitable identification framework. Building on martingale theory, the authors extend the core assumptions of discrete-time causal inference—consistency, exchangeability, and positivity—to the continuous-time setting. They formulate a dynamic treatment strategy and a potential outcomes model tailored to marked point processes and establish corresponding causal identification conditions. Leveraging this foundation, they derive a novel marginal g-formula that enables nonparametric identification of causal effects. The proposed framework subsumes existing results for discrete-time and counting process settings as special cases, demonstrating both theoretical compatibility and extensibility, thereby unifying survival analysis and causal inference within a coherent paradigm.