Large language models rely on inference-time context including rationales and instructions to improve task performance. Prior work attempts to internalize such benefits via parameter-efficient fine-tuning, and also studies inference-time representation interventions, but these approaches typically update the base model or learn fixed directions tied to global attributes. As a result, it remains unclear how to capture and reuse the internal effect of a helpful context that is specific to a given query. We take a representation-level view and define a context imprint as the hidden-state displacement between context-conditioned and query-only forward passes. We propose HintSteer, which trains a lightweight query-conditioned predictor to estimate the imprint from the query representation and injects it into a frozen model at inference time, approximating the missing context effect without explicitly providing context. Experiments on mathematical reasoning and writing demonstrate the effectiveness of imprint-based steering under no-context inference and indicate that both imprint magnitude and predictability influence the resulting gains.

Reinforcement learning with verifiable rewards (RLVR) has become a widely used approach for improving the mathematical reasoning ability of large language models (LLMs). However, recent studies show that LLMs may solve a math problem in its original form but fail when the same problem is reworded without changing its meaning. This suggests that the model has not fully acquired the underlying reasoning ability and instead relies on the surface form of the problem. In this paper, we propose PEARL, an RLVR framework that uses different versions of the same math problem as structured training signals. PEARL first groups each original problem with its checked variants to build semantic perturbation clusters, and then uses partially solved clusters as the active RL training set. PEARL further performs cluster-level RL optimization by comparing model outputs across different versions of the same mathematical problem. Experiments on mathematical reasoning benchmarks show that PEARL achieves strong results against competitive RLVR baselines while using a small active training set. These results suggest that semantically equivalent variants can serve as useful training units for improving mathematical reasoning in LLMs.

World models are emerging as critical components for assessing the consequences of actions taken by interactive agents in online planning and offline evaluation. In text-based environments, world models are typically trained and evaluated with single-step metrics such as Exact Match, aiming to improve the similarity between predicted and real-world states, but such metrics have been shown to be insufficient for capturing actual agent behavior. To address this issue, we introduce a new behavior-consistency training paradigm aimed at improving the functional consistency between the world model and the real environment. This paradigm focuses on optimizing a tractable step-level reward named Behavior Consistency Reward (BehR), which measures how much the likelihood of a logged next action changes between the real state and the world-model-predicted state under a frozen Reference Agent. Experiments on WebShop and TextWorld show that BehR-based training improves task-level functional consistency in several settings, with the clearest gains in WebShop and smaller changes in near-ceiling regimes, while preserving or improving single-step prediction quality in three of four settings. BehR-trained world models also reduce false positives in offline surrogate evaluation and improve inference-time lookahead planning over the corresponding supervised world-model baselines. Code: https://github.com/Ricardo-H/behr-wm.

Reinforcement learning with verifiable or outcome-level rewards has recently become an effective post-training paradigm for improving the reasoning abilities of large language models, with GRPO-style optimization achieving strong results on mathematical and other verifiable tasks. However, for multi-step reasoning, outcome-level supervision remains intrinsically sparse: a trajectory that contains partially correct intermediate derivations but makes a late mistake is still treated as a complete failure. As a result, useful reasoning prefixes are penalized together with erroneous continuations, and informative failed trajectories are underutilized. PRMs provide denser step-level feedback, but they often rely on costly process annotation or auxiliary process verifiers, and optimizing local process rewards alone may fail to preserve the global objective of final correctness, leaving room for reward misspecification and reward hacking during reinforcement fine-tuning. We propose PRSA(Process Rewarding via Solution-Space Alignment), a simple framework that bridges outcome supervision and process supervision without requiring human step-level labels. PRSA automatically extracts intermediate-value steps from model outputs and reference solutions, aligns them within the solution space under equivalence and order constraints to estimate reasoning progress, and uses this signal for both reward shaping and token-level prefix-suffix credit assignment. In this way, PRSA preserves the global anchor of final-answer correctness while providing denser and better localized learning signals for partially correct trajectories.

Reinforcement learning for open-ended text generation is constrained by the lack of verifiable rewards, necessitating reliance on judge models that require either annotated data or powerful closed-source models. Inspired by recent work on unsupervised reinforcement learning for mathematical reasoning using confidence-based endogenous rewards, we investigate whether this principle can be adapted to open-ended writing tasks. We find that directly applying confidence rewards leads to Triviality Bias: the policy collapses toward high-probability outputs, reducing diversity and meaningful content. We propose TCER (Triviality Corrected Endogenous Reward), which addresses this bias by rewarding the relative information gain between a specialist policy and a generalist reference policy, modulated by a probability-dependent correction mechanism. Across multiple writing benchmarks and model architectures, TCER achieves consistent improvements without external supervision. Furthermore, TCER also transfers effectively to mathematical reasoning, validating the generality of our approach across different generation tasks.

Although the effectiveness of Large Language Models as judges has been validated, their performance remains limited in open-ended tasks, particularly in story evaluation. Accurate story evaluation is crucial not only for assisting human quality judgment but also for providing reward signals to guide story generation. However, existing methods face a dilemma: prompt engineering for closed-source models suffers from poor adaptability, while fine-tuning approaches for open-source models lack the reasoning capabilities essential for story evaluation. To address this, we propose the Self-Evolving Pairwise Reasoning (EvolvR) framework. Grounded in pairwise comparison, the framework first self-synthesizes score-aligned Chain-of-Thought (CoT) data via a multi-persona strategy. To ensure data quality, these raw CoTs undergo a self-filtering process, utilizing multi-agents to guarantee their logical rigor and robustness. Finally, the evaluator trained on the refined data is deployed as a reward model to guide the story generation task. Experimental results demonstrate that our framework achieves state-of-the-art performance on three evaluation benchmarks including StoryER, HANNA and OpenMEVA. Furthermore, when served as a reward model, it enhances the quality of generated stories, thereby validating the superiority of our self-evolving approach.

Automatic prompt optimization is a practical alternative to fine-tuning for adapting large language models (LLMs), yet existing approaches often trade off signal quality against computational cost. Methods that rely on generative feedback can be informative but expensive to scale, while sampling-based optimization typically requires many evaluations and exhibits high variance. Even loss-driven prompt optimization remains limited by costly segment attribution that scales with prompt length and by overfitting to a single evaluator, which weakens transfer across model families and domains. We propose Gradient-guided Multi-judge Prompt Optimization (GMPO), a scalable framework that improves both efficiency and robustness. GMPO uses a first-order gradient approximation to score segment importance in a continuous masking direction, requiring only one forward and one backward pass. GMPO further employs a generate multi-judge design in which candidate prompt edits are proposed by a generator and selected using cross-entropy losses aggregated from multiple lightweight judge models, reducing evaluator bias and improving generalization. Experiments across math, reasoning, instruction-following evaluation, and safety robustness benchmarks demonstrate consistent gains with substantially lower optimization overhead.

Prompt optimization is a practical and widely applicable alternative to fine tuning for improving large language model performance. Yet many existing methods evaluate candidate prompts by sampling full outputs, often coupled with self critique or human annotated preferences, which limits scalability, especially for smaller models or models that are not instruction tuned. We present PMPO (Probabilistic Metric Prompt Optimization), a unified framework that uses token level cross entropy as a direct, lightweight evaluation signal. PMPO locates low quality prompt segments via a masking based analysis and iteratively rewrites them to propose improved variants. Crucially, during evaluation, PMPO selects among variants by minimizing loss in a single forward pass, eliminating output sampling and human or judge based scoring for selection while still using standard generation only to propose rewrites. This unified, loss based strategy supports both supervised and preference based tasks. Across model sizes and datasets, PMPO outperforms prior prompt optimizers: it achieves the highest average accuracy on BBH, performs strongly on GSM8K and AQuA RAT, and raises AlpacaEval 2.0 win rates by over 19 points. These results demonstrate PMPO’s effectiveness, efficiency, and broad applicability.