What If Consensus Lies? Selective-Complementary Reinforcement Learning at Test Time

Test-Time Reinforcement Learning (TTRL) enables Large Language Models (LLMs) to enhance reasoning capabilities on unlabeled test streams by deriving pseudo-rewards from majority voting consensus. However, existing TTRL methods rely exclusively on positive pseudo-labeling strategies. Such reliance becomes vulnerable under challenging scenarios where answer distributions are highly dispersed, resulting in weak consensus that inadvertently reinforces incorrect trajectories as supervision signals. In this paper, we propose SCRL (Selective-Complementary Reinforcement Learning), a robust test-time reinforcement learning framework that effectively mitigates label noise amplification. SCRL develops Selective Positive Pseudo-Labeling, which enforces strict consensus criteria to filter unreliable majorities. Complementarily, SCRL introduces Entropy-Gated Negative Pseudo-Labeling, the first negative supervision mechanism in TTRL, to reliably prune incorrect trajectories based on generation uncertainty. Extensive experiments on multiple reasoning benchmarks demonstrate that SCRL achieves substantial improvements over baselines, while maintaining robust generalization and training stability under constrained rollout budgets. Our code is available at https://github.com/Jasper-Yan/SCRL.

Taming Momentum: Rethinking Optimizer States Through Low-Rank Approximation

Modern optimizers like Adam and Muon are central to training large language models, but their reliance on first- and second-order momenta introduces significant memory overhead, which constrains scalability and computational efficiency. In this work, we reframe the exponential moving average (EMA) used in these momenta as the training of a linear regressor via online gradient flow. Building on this equivalence, we introduce LoRA-Pre, a novel low-rank optimizer designed for efficient pre-training. Specifically, LoRA-Pre reduces the optimizer's memory footprint by decomposing the full momentum matrix into a compact low-rank subspace within the online linear learner, thereby maintaining optimization performance while improving memory efficiency. We empirically validate LoRA-Pre's efficacy by pre-training models from the Llama architecture family, scaling from 60M to 1B parameters. LoRA-Pre achieves the highest performance across all model sizes. Notably, LoRA-Pre demonstrates remarkable rank efficiency, achieving comparable or superior results using only 1/8 the rank of baseline methods. Beyond pre-training, we evaluate LoRA-Pre's effectiveness in fine-tuning scenarios. With the same rank, LoRA-Pre consistently outperforms all efficient fine-tuning baselines. Specifically, compared to standard LoRA, LoRA-Pre achieves substantial improvements of 3.14 points on Llama-3.1-8B and 6.17 points on Llama-2-7B, validating our approach's effectiveness across both pre-training and fine-tuning paradigms. Our code is publicly available at https://github.com/mrflogs/LoRA-Pre.

How to Train Your Deep Research Agent? Prompt, Reward, and Policy Optimization in Search-R1

Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. While reinforcement learning (RL) has been shown to improve performance in this paradigm, its contributions remain underexplored. To fully understand the role of RL, we conduct a systematic study along three decoupled dimensions: prompt template, reward function, and policy optimization. Our study reveals that: 1) the Fast Thinking template yields greater stability and better performance than the Slow Thinking template used in prior work; 2) the F1-based reward underperforms the EM due to training collapse driven by answer avoidance; this can be mitigated by incorporating action-level penalties, ultimately surpassing EM; 3) REINFORCE outperforms PPO while requiring fewer search actions, whereas GRPO shows the poorest stability among policy optimization methods. Building on these insights, we then introduce Search-R1++, a strong baseline that improves the performance of Search-R1 from 0.403 to 0.442 (Qwen2.5-7B) and 0.289 to 0.331 (Qwen2.5-3B). We hope that our findings can pave the way for more principled and reliable RL training strategies in Deep Research systems.

Mitigating the Safety-utility Trade-off in LLM Alignment via Adaptive Safe Context Learning

While reasoning models have achieved remarkable success in complex reasoning tasks, their increasing power necessitates stringent safety measures. For safety alignment, the core challenge lies in the inherent trade-off between safety and utility. However, prevailing alignment strategies typically construct CoT training data with explicit safety rules via context distillation. This approach inadvertently limits reasoning capabilities by creating a rigid association between rule memorization and refusal. To mitigate the safety-utility trade-off, we propose the Adaptive Safe Context Learning (ASCL) framework to improve the reasoning given proper context. ASCL formulates safety alignment as a multi-turn tool-use process, empowering the model to autonomously decide when to consult safety rules and how to generate the ongoing reasoning. Furthermore, to counteract the preference for rule consultation during RL, we introduce Inverse Frequency Policy Optimization (IFPO) to rebalance advantage estimates. By decoupling rule retrieval and subsequent reasoning, our method achieves higher overall performance compared to baselines.

Stop Tracking Me! Proactive Defense Against Attribute Inference Attack in LLMs

Recent studies have shown that large language models (LLMs) can infer private user attributes (e.g., age, location, gender) from user-generated text shared online, enabling rapid and large-scale privacy breaches. Existing anonymization-based defenses are coarse-grained, lacking word-level precision in anonymizing privacy-leaking elements. Moreover, they are inherently limited as altering user text to hide sensitive cues still allows attribute inference to occur through models' reasoning capabilities. To address these limitations, we propose a unified defense framework that combines fine-grained anonymization (TRACE) with inference-preventing optimization (RPS). TRACE leverages attention mechanisms and inference chain generation to identify and anonymize privacy-leaking textual elements, while RPS employs a lightweight two-stage optimization strategy to induce model rejection behaviors, thereby preventing attribute inference. Evaluations across diverse LLMs show that TRACE-RPS reduces attribute inference accuracy from around 50\% to below 5\% on open-source models. In addition, our approach offers strong cross-model generalization, prompt-variation robustness, and utility-privacy tradeoffs. Our code is available at https://github.com/Jasper-Yan/TRACE-RPS.

Do MLLMs Really Understand Space? A Mathematical Reasoning Evaluation

Multimodal large language models (MLLMs) have achieved strong performance on perception-oriented tasks, yet their ability to perform mathematical spatial reasoning, defined as the capacity to parse and manipulate two- and three-dimensional relations, remains unclear. Humans easily solve textbook-style spatial reasoning problems with over 95\% accuracy, but we find that most leading MLLMs fail to reach even 60\% on the same tasks. This striking gap highlights spatial reasoning as a fundamental weakness of current models. To investigate this gap, we present MathSpatial, a unified framework for evaluating and improving spatial reasoning in MLLMs. MathSpatial includes three complementary components: (i) MathSpatial-Bench, a benchmark of 2K problems across three categories and eleven subtypes, designed to isolate reasoning difficulty from perceptual noise; (ii) MathSpatial-Corpus, a training dataset of 8K additional problems with verified solutions; and (iii) MathSpatial-SRT, which models reasoning as structured traces composed of three atomic operations--Correlate, Constrain, and Infer. Experiments show that fine-tuning Qwen2.5-VL-7B on MathSpatial achieves competitive accuracy while reducing tokens by 25\%. MathSpatial provides the first large-scale resource that disentangles perception from reasoning, enabling precise measurement and comprehensive understanding of mathematical spatial reasoning in MLLMs.

One Size, Many Fits: Aligning Diverse Group-Wise Click Preferences in Large-Scale Advertising Image Generation

Advertising image generation has increasingly focused on online metrics like Click-Through Rate (CTR), yet existing approaches adopt a ``one-size-fits-all" strategy that optimizes for overall CTR while neglecting preference diversity among user groups. This leads to suboptimal performance for specific groups, limiting targeted marketing effectiveness. To bridge this gap, we present \textit{One Size, Many Fits} (OSMF), a unified framework that aligns diverse group-wise click preferences in large-scale advertising image generation. OSMF begins with product-aware adaptive grouping, which dynamically organizes users based on their attributes and product characteristics, representing each group with rich collective preference features. Building on these groups, preference-conditioned image generation employs a Group-aware Multimodal Large Language Model (G-MLLM) to generate tailored images for each group. The G-MLLM is pre-trained to simultaneously comprehend group features and generate advertising images. Subsequently, we fine-tune the G-MLLM using our proposed Group-DPO for group-wise preference alignment, which effectively enhances each group's CTR on the generated images. To further advance this field, we introduce the Grouped Advertising Image Preference Dataset (GAIP), the first large-scale public dataset of group-wise image preferences, including around 600K groups built from 40M users. Extensive experiments demonstrate that our framework achieves the state-of-the-art performance in both offline and online settings. Our code and datasets will be released at https://github.com/JD-GenX/OSMF.

Learning Fair Domain Adaptation with Virtual Label Distribution

Unsupervised Domain Adaptation (UDA) aims to mitigate performance degradation when training and testing data are sampled from different distributions. While significant progress has been made in enhancing overall accuracy, most existing methods overlook performance disparities across categories-an issue we refer to as category fairness. Our empirical analysis reveals that UDA classifiers tend to favor certain easy categories while neglecting difficult ones. To address this, we propose Virtual Label-distribution-aware Learning (VILL), a simple yet effective framework designed to improve worst-case performance while preserving high overall accuracy. The core of VILL is an adaptive re-weighting strategy that amplifies the influence of hard-to-classify categories. Furthermore, we introduce a KL-divergence-based re-balancing strategy, which explicitly adjusts decision boundaries to enhance category fairness. Experiments on commonly used datasets demonstrate that VILL can be seamlessly integrated as a plug-and-play module into existing UDA methods, significantly improving category fairness.

TAT: Task-Adaptive Transformer for All-in-One Medical Image Restoration

Medical image restoration (MedIR) aims to recover high-quality medical images from their low-quality counterparts. Recent advancements in MedIR have focused on All-in-One models capable of simultaneously addressing multiple different MedIR tasks. However, due to significant differences in both modality and degradation types, using a shared model for these diverse tasks requires careful consideration of two critical inter-task relationships: task interference, which occurs when conflicting gradient update directions arise across tasks on the same parameter, and task imbalance, which refers to uneven optimization caused by varying learning difficulties inherent to each task. To address these challenges, we propose a task-adaptive Transformer (TAT), a novel framework that dynamically adapts to different tasks through two key innovations. First, a task-adaptive weight generation strategy is introduced to mitigate task interference by generating task-specific weight parameters for each task, thereby eliminating potential gradient conflicts on shared weight parameters. Second, a task-adaptive loss balancing strategy is introduced to dynamically adjust loss weights based on task-specific learning difficulties, preventing task domination or undertraining. Extensive experiments demonstrate that our proposed TAT achieves state-of-the-art performance in three MedIR tasks--PET synthesis, CT denoising, and MRI super-resolution--both in task-specific and All-in-One settings. Code is available at https://github.com/Yaziwel/TAT.

Reassessing the Role of Supervised Fine-Tuning: An Empirical Study in VLM Reasoning

Recent advances in vision-language models (VLMs) reasoning have been largely attributed to the rise of reinforcement Learning (RL), which has shifted the community's focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM Reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and strongly influenced by model capacity, data scale, and data distribution. Contrary to common assumptions, our findings show that SFT plays a crucial role across several scenarios: (1) Effectiveness for weaker models. SFT more reliably elicits reasoning capabilities in smaller or weaker VLMs. (2) Data efficiency. SFT with only 2K achieves comparable or better reasoning performance to RL with 20K. (3) Cross-modal transferability. SFT demonstrates stronger generalization across modalities. Moreover, we identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL. These results challenge the prevailing "RL over SFT" narrative. They highlight that the role of SFT may have been underestimated and support a more balanced post-training pipeline in which SFT and RL function as complementary components.