Wenjing Yan

Zhiyuan Zhai, Wenjing Yan, Xiaodan Shao et al.

Does reinforcement learning genuinely expand what LLM agents can do, or merely make them more reliable? For static reasoning, recent work answers the second: base and RL pass@k curves converge at large k. We ask whether this holds for agentic tool use, where T rounds of interaction enable compositional strategies that re-sampling cannot recover. We introduce PASS@(k,T), a two-dimensional metric that jointly varies sampling budget k and interaction depth T, separating capability expansion from efficiency improvement. Our main finding is that, contrary to the static-reasoning result, tool-use RL genuinely enlarges the capability boundary: the RL agent's pass-curve pulls above the base model's and the gap widens at large k rather than converging. The expansion is specific to compositional, sequential information gathering; on simpler tasks RL behaves as prior work predicts. Under matched training data, supervised fine-tuning regresses the boundary on the same compositional tasks, isolating self-directed exploration as the causal factor. Mechanism analysis shows RL reweights the base strategy distribution toward the subset whose downstream reasoning more often yields a correct answer, with the improvement concentrated on how the agent integrates retrieved information. These results reconcile optimistic and pessimistic readings of RL for LLMs: both are correct, on different task types.

Wenjing Yan, Xuanyu Cao

Performative prediction is a recently proposed framework where predictions guide decision-making and hence influence future data distributions. Such performative phenomena are ubiquitous in various areas, such as transportation, finance, public policy, and recommendation systems. To date, work on performative prediction has only focused on unconstrained scenarios, neglecting the fact that many real-world learning problems are subject to constraints. This paper bridges this gap by studying performative prediction under inequality constraints. Unlike most existing work that provides only performative stable points, we aim to find the optimal solutions. Anticipating performative gradients is a challenging task, due to the agnostic performative effect on data distributions. To address this issue, we first develop a robust primal-dual framework that requires only approximate gradients up to a certain accuracy, yet delivers the same order of performance as the stochastic primal-dual algorithm without performativity. Based on this framework, we then propose an adaptive primal-dual algorithm for location families. Our analysis demonstrates that the proposed adaptive primal-dual algorithm attains $\ca{O}(\sqrt{T})$ regret and constraint violations, using only $\sqrt{T} + 2T$ samples, where $T$ is the time horizon. To our best knowledge, this is the first study and analysis on the optimality of the performative prediction problem under inequality constraints. Finally, we validate the effectiveness of our algorithm and theoretical results through numerical simulations.

Zhiyuan Zhai, Xinkai You, Wenjing Yan et al.

Reasoning-capable large language models solve hard problems by emitting long chains of thought, paying heavily in latency, GPU time, and energy. Casual inspection of their traces reveals extensive reformulation, verification, and circular self-reflection, yet how much of this deliberation is actually necessary has never been measured at scale or explained from first principles. This paper closes both gaps. We formalise reasoning redundancy directly in terms of the reasoning model itself: the redundancy of a correct trace is the largest fraction of its trailing segmented steps that can be truncated while $π$, forced to terminate thinking and emit a final answer, still produces the correct answer. A large-scale quantification across four frontier reasoning models and two mathematical benchmarks shows that step-level redundancy is consistently high -- between 61% and 93% across the 8 (model, benchmark) conditions we study, with the median critical prefix equal to a single segmented step in six of the eight conditions -- that the finding is robust to the choice of judge family, and that although $ρ$ decreases with problem difficulty on MATH-500, all four models remain substantially redundant ($ρ\in [46\%, 85\%]$) even on the hardest Level-5 problems. We then prove that this redundancy is a structural consequence of length-agnostic outcome rewards, not a model-specific artefact: under any such reward, no finite expected stopping time is optimal. The result holds regardless of RL algorithm, base model, data distribution, or whether the policy is obtained via RL or distillation; over-thinking is therefore not a bug to be patched in individual models but a structural property of how current reasoning models are trained. Code: https://github.com/zhiyuanZhai20/how-much-thinking-is-enough

Wenjing Yan, Xuanyu Cao

Decentralized optimization methods often entail information exchange between neighbors. Transmission failures can happen due to network congestion, hardware/software issues, communication outage, and other factors. In this paper, we investigate the random link failure problem in decentralized multi-task online convex optimization, where agents have individual decisions that are coupled with each other via pairwise constraints. Although widely used in constrained optimization, conventional saddle-point algorithms are not directly applicable here because of random packet dropping. To address this issue, we develop a robust decentralized saddle-point algorithm against random link failures with heterogeneous probabilities by replacing the missing decisions of neighbors with their latest received values. Then, by judiciously bounding the accumulated deviation stemming from this replacement, we first establish that our algorithm achieves $\mathcal{O}(\sqrt{T})$ regret and $\mathcal{O}(T^\frac{3}{4})$ constraint violations for the full information scenario, where the complete information on the local cost function is revealed to each agent at the end of each time slot. These two bounds match, in order sense, the performance bounds of algorithms with perfect communications. Further, we extend our algorithm and analysis to the two-point bandit feedback scenario, where only the values of the local cost function at two random points are disclosed to each agent sequentially. Performance bounds of the same orders as the full information case are derived. Finally, we corroborate the efficacy of the proposed algorithms and the analytical results through numerical simulations.

Hongsong Wang, Wenjing Yan, Qiuxia Lai et al.

Text-to-Motion (T2M) generation aims to synthesize realistic human motion sequences from natural language descriptions. While two-stage frameworks leveraging discrete motion representations have advanced T2M research, they often neglect cross-sequence temporal consistency, i.e., the shared temporal structures present across different instances of the same action. This leads to semantic misalignments and physically implausible motions. To address this limitation, we propose TCA-T2M, a framework for temporal consistency-aware T2M generation. Our approach introduces a temporal consistency-aware spatial VQ-VAE (TCaS-VQ-VAE) for cross-sequence temporal alignment, coupled with a masked motion transformer for text-conditioned motion generation. Additionally, a kinematic constraint block mitigates discretization artifacts to ensure physical plausibility. Experiments on HumanML3D and KIT-ML benchmarks demonstrate that TCA-T2M achieves state-of-the-art performance, highlighting the importance of temporal consistency in robust and coherent T2M generation.

Chenrui Xu, Wenjing Yan, Ying-Jun Angela Zhang

Training large-scale neural networks requires solving nonconvex optimization where the choice of optimizer fundamentally determines both convergence behavior and computational efficiency. While adaptive methods like Adam have long dominated practice, the recently proposed Muon optimizer achieves superior performance through orthogonalized momentum updates that enforce isotropic geometry with uniform singular values. However, this strict isotropy discards potentially valuable curvature information encoded in gradient spectra, motivating optimization methods that balance geometric structure with adaptivity. We introduce FISMO (Fisher-Structured Momentum-Orthogonalized) optimizer, which generalizes isotropic updates to incorporate anisotropic curvature information through Fisher information geometry. By reformulating the optimizer update as a trust-region problem constrained by a Kronecker-factored Fisher metric, FISMO achieves structured preconditioning that adapts to local loss landscape geometry while maintaining computational tractability. We establish convergence guarantees for FISMO in stochastic nonconvex settings, proving an $\mathcal{O}(1/\sqrt{T})$ rate for the expected squared gradient norm with explicit characterization of variance reduction through mini-batching. Empirical evaluation on image classification and language modeling benchmarks demonstrates that FISMO achieves superior training efficiency and final performance compared to established baselines.

Ruichen Xu, Wenjing Yan, Ying-Jun Angela Zhang

Understanding reasoning in large language models is complicated by evaluations that conflate multiple reasoning types. We isolate analogical reasoning (inferring shared properties between entities based on known similarities) and analyze its emergence in transformers. We theoretically prove three key results: (1) Joint training on similarity and attribution premises enables analogical reasoning through aligned representations; (2) Sequential training succeeds only when similarity structure is learned before specific attributes, revealing a necessary curriculum; (3) Two-hop reasoning ($a \to b, b \to c \implies a \to c$) reduces to analogical reasoning with identity bridges ($b = b$), which must appear explicitly in training data. These results reveal a unified mechanism: transformers encode entities with similar properties into similar representations, enabling property transfer through feature alignment. Experiments with architectures up to 1.5B parameters validate our theory and demonstrate how representational geometry shapes inductive reasoning capabilities.

Zhiwei Zhai, Wenjing Yan, Ying-Jun Angela Zhang

Decentralized bilevel optimization has garnered significant attention due to its critical role in solving large-scale machine learning problems. However, existing methods often rely on prior knowledge of problem parameters-such as smoothness, convexity, or communication network topologies-to determine appropriate stepsizes. In practice, these problem parameters are typically unavailable, leading to substantial manual effort for hyperparameter tuning. In this paper, we propose AdaSDBO, a fully problem-parameter-free algorithm for decentralized bilevel optimization with a single-loop structure. AdaSDBO leverages adaptive stepsizes based on cumulative gradient norms to update all variables simultaneously, dynamically adjusting its progress and eliminating the need for problem-specific hyperparameter tuning. Through rigorous theoretical analysis, we establish that AdaSDBO achieves a convergence rate of $\widetilde{\mathcal{O}}\left(\frac{1}{T}\right)$, matching the performance of well-tuned state-of-the-art methods up to polylogarithmic factors. Extensive numerical experiments demonstrate that AdaSDBO delivers competitive performance compared to existing decentralized bilevel optimization methods while exhibiting remarkable robustness across diverse stepsize configurations.

Wenjing Yan, Shan Li, Chengtao Que et al.

Much of the work on automatic facial expression recognition relies on databases containing a certain number of emotion classes and their exaggerated facial configurations (generally six prototypical facial expressions), based on Ekman's Basic Emotion Theory. However, recent studies have revealed that facial expressions in our human life can be blended with multiple basic emotions. And the emotion labels for these in-the-wild facial expressions cannot easily be annotated solely on pre-defined AU patterns. How to analyze the action units for such complex expressions is still an open question. To address this issue, we develop a RAF-AU database that employs a sign-based (i.e., AUs) and judgement-based (i.e., perceived emotion) approach to annotating blended facial expressions in the wild. We first reviewed the annotation methods in existing databases and identified crowdsourcing as a promising strategy for labeling in-the-wild facial expressions. Then, RAF-AU was finely annotated by experienced coders, on which we also conducted a preliminary investigation of which key AUs contribute most to a perceived emotion, and the relationship between AUs and facial expressions. Finally, we provided a baseline for AU recognition in RAF-AU using popular features and multi-label learning methods.