SAGA: Scene-Aware, Goal-Evolving Agents for Long-Horizon CivRealm Strategy Planning

Long-horizon strategic planning in complex strategy games demands concurrent reasoning across multiple decision domains under imperfect information and sparse reward. Existing LLM-based agents suffer from three systematic failures: scene blindness from raw tile coordinates, context overflow and domain coupling from monolithic state dumps, and shallow cross-game learning that treats each episode in isolation. We present SAGA, an LLM multi-agent framework with three mechanisms each directly targeting one class of failure: (i) a Map-Semantic Scene Graph that encodes typed spatial relations among game entities into per-unit natural-language context, resolving spatial blindness without global token inflation; (ii) a Tool-Augmented Planner that pulls fine-grained domain state on demand and dispatches per-domain directives to dedicated specialist controllers, eliminating context overflow, domain coupling, and mechanical constraint violations; and (iii) a Dual-Horizon Feedback Loop that combines periodic within-game goal generation with structured cross-game causal post-mortem, enabling principled strategic evolution without manual reward engineering. Evaluated on FreeCiv, SAGA attains the highest mean civilization score -- the environment's sole sparse objective reward -- with lower variance than the two strongest baselines, and is the only method that significantly surpasses every baseline on infrastructure construction, the resource axis most readily sacrificed under multi-objective conflict. It outscores the two strongest baselines in most head-to-head games while cutting output tokens (the dominant decoding cost) by 27%. Equipped with the cross-game evolution module, SAGA reaches the highest end-of-chain score across five successive episodes. Ablation studies confirm that each architectural component contributes independently to this advantage.

Black-Box Continual Learning for Vision-Language Models

The rapid deployment of Vision-Language Models (VLMs) in dynamic environments necessitates the ability to learn continuously without forgetting. However, traditional continual learning (CL) settings often rely on white-box paradigms, which is increasingly invalidated by the shift toward cloud-hosted models. In this paper, we introduce Black-CL, a more realistic benchmark for VLMs that enforces three primary real-world challenges: weight and architecture inaccessibility, constrained computation, and task-agnostic inference. The learner can query only output embeddings or logits, with no gradient flow through or structural modification of the backbone. Current CL methodologies, which rely on backbone backpropagation or complex parameter expansion, are fundamentally incompatible with these constraints. Under this setting, we propose BETA, a simple yet effective baseline built on the key insight that solely optimizing textual prototypes can navigate the complexities of CL. BETA integrates three core components: Semantic Projection Accumulation (SPA) for incremental knowledge acquisition, Latent Distribution Replay (LDR) for anchoring the embedding space against catastrophic forgetting, and Test-Time Prototype Adaptation (TTPA) for dynamic, instance-aware boundary refinement. Extensive experiments across ten diverse datasets and various backbones demonstrate that BETA significantly outperforms existing black-box tuners. Remarkably, with only 0.05 M trainable parameters, a 180--3000$\times$ reduction compared to competitive methods, BETA achieves performance on par with or even exceeding white-box CL methods. We believe Black-CL and BETA provide a foundational framework for future advancements in continual learning and accelerates the transition of continual learning from academia to real-world systems.

Large Language Models Explore by Latent Distilling

Generating diverse responses is crucial for test-time scaling of large language models (LLMs), yet standard stochastic sampling mostly yields surface-level lexical variation, limiting semantic exploration. In this paper, we propose Exploratory Sampling (ESamp), a decoding approach that explicitly encourages semantic diversity during generation. ESamp is motivated by the well-known observation that neural networks tend to make lower-error predictions on inputs similar to those encountered before, and incur higher prediction error on novel ones. Building on this property, we train a lightweight Distiller at test time to predict deep-layer hidden representations of the LLM from its shallow-layer representations to model the LLM's depth-wise representation transitions. During decoding, the Distiller continuously adapts to the mappings induced by the current generation context. ESamp uses the prediction error as a novelty signal to reweight candidate token extensions conditioned on the current prefix, thereby biasing decoding toward less-explored semantic patterns. ESamp is implemented with an asynchronous training--inference pipeline, with less than 5% worst case overhead (1.2% in the optimized release). Empirical results show that ESamp significantly boosts the Pass@k efficiency of reasoning models, showing superior or comparable performance to strong stochastic and heuristic baselines. Notably, ESamp achieves robust generalization across mathematics, science, and code generation benchmarks and breaks the trade-off between diversity and coherence in creative writing. Our code has released at: https://github.com/LinesHogan/tLLM.

IDER: IDempotent Experience Replay for Reliable Continual Learning

Catastrophic forgetting, the tendency of neural networks to forget previously learned knowledge when learning new tasks, has been a major challenge in continual learning (CL). To tackle this challenge, CL methods have been proposed and shown to reduce forgetting. Furthermore, CL models deployed in mission-critical settings can benefit from uncertainty awareness by calibrating their predictions to reliably assess their confidences. However, existing uncertainty-aware continual learning methods suffer from high computational overhead and incompatibility with mainstream replay methods. To address this, we propose idempotent experience replay (IDER), a novel approach based on the idempotent property where repeated function applications yield the same output. Specifically, we first adapt the training loss to make model idempotent on current data streams. In addition, we introduce an idempotence distillation loss. We feed the output of the current model back into the old checkpoint and then minimize the distance between this reprocessed output and the original output of the current model. This yields a simple and effective new baseline for building reliable continual learners, which can be seamlessly integrated with other CL approaches. Extensive experiments on different CL benchmarks demonstrate that IDER consistently improves prediction reliability while simultaneously boosting accuracy and reducing forgetting. Our results suggest the potential of idempotence as a promising principle for deploying efficient and trustworthy continual learning systems in real-world applications.Our code is available at https://github.com/YutingLi0606/Idempotent-Continual-Learning.

Adaptive Value Decomposition: Coordinating a Varying Number of Agents in Urban Systems

Multi-agent reinforcement learning (MARL) provides a promising paradigm for coordinating multi-agent systems (MAS). However, most existing methods rely on restrictive assumptions, such as a fixed number of agents and fully synchronous action execution. These assumptions are often violated in urban systems, where the number of active agents varies over time, and actions may have heterogeneous durations, resulting in a semi-MARL setting. Moreover, while sharing policy parameters among agents is commonly adopted to improve learning efficiency, it can lead to highly homogeneous actions when a subset of agents make decisions concurrently under similar observations, potentially degrading coordination quality. To address these challenges, we propose Adaptive Value Decomposition (AVD), a cooperative MARL framework that adapts to a dynamically changing agent population. AVD further incorporates a lightweight mechanism to mitigate action homogenization induced by shared policies, thereby encouraging behavioral diversity and maintaining effective cooperation among agents. In addition, we design a training-execution strategy tailored to the semi-MARL setting that accommodates asynchronous decision-making when some agents act at different times. Experiments on real-world bike-sharing redistribution tasks in two major cities, London and Washington, D.C., demonstrate that AVD outperforms state-of-the-art baselines, confirming its effectiveness and generalizability.

Spectral Disentanglement and Enhancement: A Dual-domain Contrastive Framework for Representation Learning

Large-scale multimodal contrastive learning has recently achieved impressive success in learning rich and transferable representations, yet it remains fundamentally limited by the uniform treatment of feature dimensions and the neglect of the intrinsic spectral structure of the learned features. Empirical evidence indicates that high-dimensional embeddings tend to collapse into narrow cones, concentrating task-relevant semantics in a small subspace, while the majority of dimensions remain occupied by noise and spurious correlations. Such spectral imbalance and entanglement undermine model generalization. We propose Spectral Disentanglement and Enhancement (SDE), a novel framework that bridges the gap between the geometry of the embedded spaces and their spectral properties. Our approach leverages singular value decomposition to adaptively partition feature dimensions into strong signals that capture task-critical semantics, weak signals that reflect ancillary correlations, and noise representing irrelevant perturbations. A curriculum-based spectral enhancement strategy is then applied, selectively amplifying informative components with theoretical guarantees on training stability. Building upon the enhanced features, we further introduce a dual-domain contrastive loss that jointly optimizes alignment in both the feature and spectral spaces, effectively integrating spectral regularization into the training process and encouraging richer, more robust representations. Extensive experiments on large-scale multimodal benchmarks demonstrate that SDE consistently improves representation robustness and generalization, outperforming state-of-the-art methods. SDE integrates seamlessly with existing contrastive pipelines, offering an effective solution for multimodal representation learning.

Grad2Reward: From Sparse Judgment to Dense Rewards for Improving Open-Ended LLM Reasoning

Reinforcement Learning with Verifiable Rewards (RLVR) has catalyzed significant breakthroughs in complex LLM reasoning within verifiable domains, such as mathematics and programming. Recent efforts have sought to extend this paradigm to open-ended tasks by employing LLMs-as-a-Judge to provide sequence-level rewards for policy optimization. However, these rewards are inherently sparse, failing to provide the fine-grained supervision necessary for generating complex, long-form trajectories. Furthermore, current work treats the Judge as a black-box oracle, discarding the rich intermediate feedback signals encoded in it. To address these limitations, we introduce Grad2Reward, a novel framework that extracts dense process rewards directly from the Judge's model inference process via a single backward pass. By leveraging gradient-based attribution, Grad2Reward enables precise token-level credit assignment, substantially enhancing training efficiency and reasoning quality. Additionally, Grad2Reward introduces a self-judging mechanism, allowing the policy to improve through its own evaluative signals without training specialized reward models or reliance on superior external Judges. The experiments demonstrate that policies optimized with Grad2Reward achieve outstanding performance across diverse open-ended tasks, affirming its effectiveness and broad generalizability.

Linking Process to Outcome: Conditional Reward Modeling for LLM Reasoning

Process Reward Models (PRMs) have emerged as a promising approach to enhance the reasoning capabilities of large language models (LLMs) by guiding their step-by-step reasoning toward a final answer. However, existing PRMs either treat each reasoning step in isolation, failing to capture inter-step dependencies, or struggle to align process rewards with the final outcome. Consequently, the reward signal fails to respect temporal causality in sequential reasoning and faces ambiguous credit assignment. These limitations make downstream models vulnerable to reward hacking and lead to suboptimal performance. In this work, we propose Conditional Reward Modeling (CRM) that frames LLM reasoning as a temporal process leading to a correct answer. The reward of each reasoning step is not only conditioned on the preceding steps but also explicitly linked to the final outcome of the reasoning trajectory. By enforcing conditional probability rules, our design captures the causal relationships among reasoning steps, with the link to the outcome allowing precise attribution of each intermediate step, thereby resolving credit assignment ambiguity. Further, through this consistent probabilistic modeling, the rewards produced by CRM enable more reliable cross-sample comparison. Experiments across Best-of-N sampling, beam search and reinforcement learning demonstrate that CRM consistently outperforms existing reward models, offering a principled framework for enhancing LLM reasoning. In particular, CRM is more robust to reward hacking and delivers stable downstream improvements without relying on verifiable rewards derived from ground truth.

Learning to Select In-Context Demonstration Preferred by Large Language Model

In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks during inference using only a few demonstrations. However, ICL performance is highly dependent on the selection of these demonstrations. Recent work explores retrieval-based methods for selecting query-specific demonstrations, but these approaches often rely on surrogate objectives such as metric learning, failing to directly optimize ICL performance. Consequently, they struggle to identify truly beneficial demonstrations. Moreover, their discriminative retrieval paradigm is ineffective when the candidate pool lacks sufficient high-quality demonstrations. To address these challenges, we propose GenICL, a novel generative preference learning framework that leverages LLM feedback to directly optimize demonstration selection for ICL. Experiments on 19 datasets across 11 task categories demonstrate that GenICL achieves superior performance than existing methods in selecting the most effective demonstrations, leading to better ICL performance.

Causal Graph Guided Steering of LLM Values via Prompts and Sparse Autoencoders

As large language models (LLMs) become increasingly integrated into critical applications, aligning their behavior with human values presents significant challenges. Current methods, such as Reinforcement Learning from Human Feedback (RLHF), often focus on a limited set of values and can be resource-intensive. Furthermore, the correlation between values has been largely overlooked and remains underutilized. Our framework addresses this limitation by mining a causal graph that elucidates the implicit relationships among various values within the LLMs. Leveraging the causal graph, we implement two lightweight mechanisms for value steering: prompt template steering and Sparse Autoencoder feature steering, and analyze the effects of altering one value dimension on others. Extensive experiments conducted on Gemma-2B-IT and Llama3-8B-IT demonstrate the effectiveness and controllability of our steering methods.