Minglai Shao

Explicit Chain-of-Thought (CoT) improves the reasoning performance of large language models (LLMs) but often incurs high inference cost due to verbose token-level traces. While recent approaches reduce this overhead via concise prompting or step pruning, they largely truncate what the model says rather than internalize what the model thinks. Latent reasoning offers a promising alternative by performing computation in the hidden space, yet prior methods face two critical challenges. Many existing approaches rely on rigid point-to-point alignment, forcing a latent token to approximate the final representation of a reasoning step, which can be insufficient to capture the dense, variable-length semantics of an entire reasoning segment. Furthermore, these methods often suffer from a lack of interpretability: latent states are commonly produced by unconstrained optimization or embedding mixing, yielding vectors that are difficult to decode or audit under the pretrained language head. We propose SPOT (Span-level Pause Of Thought), a flexible framework that compresses explicit CoT into compact latent tokens without enforcing a fixed response template. At the core of SPOT is Span-level Semantic Alignment, a Sinkhorn optimal-transport objective that softly matches each token to the semantics of an entire reasoning segment, overcoming the rigidity of step-end alignment. To further improve interpretability, SPOT introduces a Frozen-Head Decoding Constraint that keeps latent states directly decodable as token distributions under the frozen pretrained LM head, enabling readable keyword interpretations of latent thoughts. Moreover, with two-stage training and without direct token-level supervision on , SPOT enables controllable implicit reasoning intensity at inference. Experiments on main reasoning benchmarks demonstrate that SPOT improves accuracy by 2.3 points on average while reducing generated tokens by 37.5% and provides faithful semantic interpretations of the latent reasoning process.

Large Language Models (LLMs) excel at multi-step reasoning, yet increasing the structural complexity of inference does not consistently improve system-level returns. Methods such as Tree of Thoughts (ToT), Graph of Thoughts (GoT), and Adaptive Graph of Thoughts (AGoT) can boost accuracy on some benchmarks, but often introduce substantial overhead in token consumption and latency, and their gains can be unstable across task distributions-sometimes underperforming simpler Chain-of-Thought (CoT) or direct input-output prompting (IO). We attribute this inefficiency to stage-wise and node-wise heterogeneity inside GoT-style reasoning pipelines: high-quality planning and final synthesis are globally coupled and typically benefit from strong models, whereas many intermediate subtasks are localized and can be solved accurately by lighter models with far fewer tokens. Motivated by these observations, we propose RouteGoT, a budget-controllable, node-adaptive routing framework for graph-structured reasoning. RouteGoT performs data-driven in-graph routing by prioritizing strong models for planning and synthesis, while dynamically allocating lightweight models and cost-effective strategies to leaf subtasks based on learned success and budget signals. It further integrates explicit budget constraints into a global inference scheduler to control graph expansion (depth/branching) under a user-specified token budget, enabling predictable performance-cost trade-offs. Experiments across reasoning, retrieval, and multi-hop QA benchmarks show that RouteGoT matches or improves accuracy while substantially reducing token usage; specifically, it achieves an average 8.1 percentage points accuracy improvement and 79.1% output token reduction compared to AGoT. Furthermore, RouteGoT outperforms existing routing baselines by maintaining a superior cost-accuracy trade-off, demonstrating improved robustness under varying budget targets and task distributions. Our code is available at https://github.com/karmaisacat-13/RouteGoT.

Deploying Generative AI-as-a-Service (GenAIaaS) at the network edge enables ubiquitous intelligence but faces severe stability challenges, primarily stemming from the lack of cognitive awareness regarding stochastic inference latency spikes and the coordination complexity of interdependent microservice chains. To bridge these gaps under strict Quality of Service (QoS) constraints, we propose value-adaptive multi-agent proximal policy optimization (VAMAPPO) as a novel cognitive orchestration framework. Unlike centralized approaches, VAMAPPO treats orchestration as a decentralized cognitive decision process. It integrates a variational inference-based uncertainty quantification mechanism, empowering agents to perceive environmental volatility and dynamically optimize learning steps to prevent performance degradation. To further solve the ambiguity in distinguishing individual contributions within sequential service chains, we design a coordinated advantage function (CAF). By utilizing counterfactual baselines to estimate marginal impacts, CAF effectively disentangles complex inter-agent dependencies. Extensive experiments on real-world network topologies with trace-driven workloads demonstrate that our cognitive framework significantly outperforms state-of-the-art baselines. Specifically, VAMAPPO achieves up to a 97.0% service success rate and reduces response latency by 24.2% while satisfying stringent service level agreement (SLA) constraints, exhibiting superior robustness in zero-shot generalization across unseen edge environments. Code and data are available at https://github.com/gymorsiback/Value-Adaptive-Multi-Agent-PPO.

Graph incremental learning aims to sequentially adapt models to evolving graphs while mitigating catastrophic forgetting. This problem becomes particularly challenging due to the simultaneous occurrence of covariate and label distribution shifts, introduced by newly emerging node classes, changes in node feature styles, and additional edges. To address these challenges, we propose DINGLE, a novel framework for both class and domain (class-domain) incremental learning on graphs. DINGLE consists of two key modules: a representation decoupler, which disentangles node representations into domain-invariant semantic factors for classification and domain-specific variation factors, and a teacher-student knowledge distillation module, which facilitates knowledge transfer across tasks while mitigating catastrophic forgetting through memory replay. By leveraging a Representative Node Feature (RNF) bank and an Encoder Parameters (EP) bank, DINGLE ensures effective knowledge retention and adaptation. Extensive experiments on 5 real-world datasets demonstrate that DINGLE outperforms 11 state-of-the-art baselines in class-domain incremental learning, improving classification accuracy while effectively preventing forgetting across tasks.

Large language models (LLMs) are increasingly applied to sequential decision-making through in-context learning (ICL), yet their effectiveness is highly sensitive to prompt quality. Effective prompts should meet three principles: focus on decision-critical information, provide step-level granularity, and minimize reliance on expert annotations through label efficiency. However, existing ICL methods often fail to satisfy all three criteria simultaneously. Motivated by these challenges, we introduce SkillGen, a skill-based ICL framework for structured sequential reasoning. It constructs an action-centric, domain-level graph from sampled trajectories, identifies high-utility actions via temporal-difference credit assignment, and retrieves step-wise skills to generate fine-grained, context-aware prompts. We further present a theoretical analysis showing that focusing on high-utility segments supports task identifiability and informs more effective ICL prompt design. Experiments on ALFWorld, BabyAI, and ScienceWorld, using both open-source and proprietary LLMs, show that SkillGen achieves consistent gains, improving progress rate by 5.9%-16.5% on average across models.

Out-of-distribution (OOD) detection is committed to delineating the classification boundaries between in-distribution (ID) and OOD images. Recent advances in vision-language models (VLMs) have demonstrated remarkable OOD detection performance by integrating both visual and textual modalities. In this context, negative prompts are introduced to emphasize the dissimilarity between image features and prompt content. However, these prompts often include a broad range of non-ID features, which may result in suboptimal outcomes due to the capture of overlapping or misleading information. To address this issue, we propose Positive and Negative Prompt Supervision, which encourages negative prompts to capture inter-class features and transfers this semantic knowledge to the visual modality to enhance OOD detection performance. Our method begins with class-specific positive and negative prompts initialized by large language models (LLMs). These prompts are subsequently optimized, with positive prompts focusing on features within each class, while negative prompts highlight features around category boundaries. Additionally, a graph-based architecture is employed to aggregate semantic-aware supervision from the optimized prompt representations and propagate it to the visual branch, thereby enhancing the performance of the energy-based OOD detector. Extensive experiments on two benchmarks, CIFAR-100 and ImageNet-1K, across eight OOD datasets and five different LLMs, demonstrate that our method outperforms state-of-the-art baselines.