Recent breakthroughs in large language models (LLMs) on reasoning tasks rely heavily on massive, high-quality datasets-typically human-annotated and thus difficult to scale. While data synthesis or distillation offers a promising alternative, existing methods struggle with inconsistent data quality and an inability to dynamically adapt to the evolving capabilities of the model, leading to suboptimal training signals. To address these limitations, we introduce Socratic-Zero, a fully autonomous framework that generates high-quality training data from minimal seed examples through the co-evolution of three agents: the Teacher, the Solver, and the Generator. The Solver continuously refines its reasoning by learning from preference feedback on both successful and failed trajectories; the Teacher adaptively crafts increasingly challenging questions based on the Solver's weaknesses; and the Generator distills the Teacher's question-design strategy to enable scalable, high-fidelity curriculum generation. This closed-loop system produces a self-improving curriculum-requiring no pre-existing tasks or labels. Remarkably, starting from only 100 seed questions, our Socratic-Solver-8B achieves an average gain of +20.2 percentage points over prior data synthesis methods across seven mathematical reasoning benchmarks (AMC23, AIME24-25, Olympiad, MATH-500, Minerva, and GSM8K), with consistent gains on both Qwen3 and GLM4 series models. Even more surprisingly, synthetic data from Socratic-Generator-32B enables student LLMs to achieve superior performance compared to other state-of-the-art (SOTA) commercial LLMs on these benchmarks, including Qwen3-235B-A22B, DeepSeek-V3.1-671B, GPT-5, Gemini-2.5-Pro, Grok-4, and Claude-4.1-Opus.

Personalizing large language models (LLMs) to individual user preferences is a critical step beyond generating generically helpful responses. However, current personalization methods are ill-suited for new users, as they typically require either slow, resource-intensive fine-tuning or a substantial amount of pre-existing user data, creating a significant cold-start problem. To address this challenge, we introduce a new paradigm for real-time personalization by learning from online pairwise preference feedback collected during text generation. We propose T-POP (T est-Time P ersonalization with O nline P reference Feedback), a novel algorithm that synergistically combines test-time alignment with dueling bandits. Without updating the LLM parameters, T-POP steers the decoding process of a frozen LLM by learning a reward function online that captures user preferences. By leveraging dueling bandits, T-POP intelligently queries the user to efficiently balance between exploring their preferences and exploiting the learned knowledge to generate personalized text. Extensive experiments demonstrate that T-POP achieves rapid and data-efficient personalization, significantly outperforming existing baselines and showing consistent improvement with more user interactions. While large language models (LLMs) have achieved remarkable success in generating human-like text, a critical frontier remains: moving from generic, one-size-fits-all responses to deeply personalized interactions.

In this work, we introduce SPLICE, a human-curated benchmark derived from the COIN instructional video dataset, designed to probe event-based reasoning across multiple dimensions: temporal, causal, spatial, contextual, and general knowledge. SPLICE includes 3,381 human-filtered videos spanning 12 categories and 180 sub-categories, such as sports, engineering, and housework. These videos are segmented into a total of 11,423 event clips. We evaluate both human participants and state-of-the-art vision-language models (VLMs) on the task of rearranging these clips into coherent event sequences to assess visual reasoning capabilities. Results reveal a significant gap: VLMs struggle to match human performance. While human-annotated textual descriptions improve model accuracy, they do not affect human performance, suggesting that models rely more on language priors than on visual understanding. Even with annotations, VLMs fall short of human-level reasoning, underscoring persistent challenges in visual reasoning. A deeper analysis across sub-categories shows that VLMs perform relatively better on videos where temporal and causal reasoning are dominant, compared to those where contextual and spatial reasoning are dominant. They also perform better on everyday tasks than on specialized ones.

Flow map models such as Consistency Models (CM) and Mean Flow (MF) enable few-step generation by learning the long jump of the ODE solution of diffusion models, yet training remains unstable, sensitive to hyperparameters, and costly. Initializing from a pre-trained diffusion model helps, but still requires converting infinitesimal steps into a long-jump map, leaving instability unresolved. We introduce mid-training, the first concept and practical method that inserts a lightweight intermediate stage between the (diffusion) pre-training and the final flow map training (i.e., post-training) for vision generation. Concretely, Consistency Mid-Training (CMT) is a compact and principled stage that trains a model to map points along a solver trajectory from a pre-trained model, starting from a prior sample, directly to the solver-generated clean sample. It yields a trajectory-consistent and stable initialization. This initializer outperforms random and diffusion-based baselines and enables fast, robust convergence without heuristics. Initializing post-training with CMT weights further simplifies flow map learning. Empirically, CMT achieves state of the art two step FIDs: 1.97 on CIFAR-10, 1.32 on ImageNet 64x64, and 1.84 on ImageNet 512x512, while using up to 98% less training data and GPU time, compared to CMs. On ImageNet 256x256, CMT reaches 1-step FID 3.34 while cutting total training time by about 50% compared to MF from scratch (FID 3.43). This establishes CMT as a principled, efficient, and general framework for training flow map models.

Process reward models (PRMs) have demonstrated significant efficacy in enhancing the mathematical reasoning capabilities of large language models (LLMs) by leveraging test-time scaling (TTS). However, while most PRMs exhibit substantial gains in mathematical domains, the scarcity of domain-specific training data and knowledge-based learning patterns limits their generalization ability when faced with other domains. T o address this limitation, we shift the learning objective from verifying domain-specific knowledge to modeling domain-agnostic logical flow. Centering on contextual coherence between chain-of-thought (CoT) steps, our approach is realized through a novel data annotation and training framework, which enhances the model's generalization capabilities across diverse domains. F or instance, our resulting model, ContextPRM, achieves a notable 6.5% average accuracy improvement over the majority voting baseline via weighted majority voting across nine non-mathematical domains in MMLU-Pro, including law, history, and philosophy, significantly surpassing the 2.2% improvement from V ersaPRM and 0.5% gains from other mathematics-focused PRMs, demonstrating consistent performance across both mathematical and non-mathematical domains.

Recent advances in Large Language Models (LLMs) (V aswani et al., 2017; Radford et al., 2019; Naveed et al., 2025) and Multimodal Large Language Models (MLLMs) (Zhang et al., 2024a; Yin et al., 2024; Wu et al., 2023) have reinforced the paradigm of language as a universal interface, substantially improving cross-modal perception and reasoning. Extending this progress to 3D, recent research has focused on 3D-aware Multimodal Large Language Models (3D-MLLMs) (Ren et al., 2025), which unify 3D scene understanding and vision-language reasoning under a linguistic interface. This line of work underscores the importance of grounding language in persistent 3D spatial representations (Cheng et al., 2024a; Roh et al., 2022), offering a unified pathway toward systematic scene-level reasoning. Recent studies have made substantial progress in 3D vision-language (3D VL) reasoning (Chen et al., 2024c; Huang et al., 2023b), yet most approaches rely on complex 3D inputs, incurring high costs in data collection, preprocessing, and computation. Some models rely on point clouds or reconstructed scenes augmented with rendered views or semantic-geometric features (Hong et al., 2023a; Fu et al., 2024), while others adopt simpler inputs but still depend on explicit 3D scene representations such as reconstructed objects aligned with semantic representations (Chu et al., 2024; Huang et al., 2023a; 2024). Despite their effectiveness, these pipelines depend on depth, poses, or external modules, leading to substantial data and engineering overhead as well as high memory and latency costs.

Learning from procedural videos remains a core challenge in self-supervised representation learning, as real-world instructional data often contains background segments, repeated actions, and steps presented out of order. Such variability violates the strong monotonicity assumptions underlying many alignment methods. Prior state-of-the-art approaches, such as OPEL, leverage Kantorovich Optimal Transport (KOT) to build frame-to-frame correspondences, but rely solely on feature similarity and fail to capture the higher-order temporal structure of a task. In this paper, we introduce REALIGN, a self-supervised framework for procedure learning based on Regularized Fused Partial Gromov-Wasserstein Optimal Transport (R-FPGWOT). In contrast to KOT, our formulation jointly models visual correspondences and temporal relations under a partial alignment scheme, enabling robust handling of irrelevant frames, repeated actions, and non-monotonic step orders common in instructional videos. To stabilize training, we integrate FPGWOT distances with inter-sequence contrastive learning, avoiding the need for multiple regularizers and preventing collapse to degenerate solutions. Across egocentric (EgoProceL) and third-person (ProceL, CrossTask) benchmarks, REALIGN achieves up to 18.9% average F1-score improvements and over 30% temporal IoU gains, while producing more interpretable transport maps that preserve key-step orderings and filter out noise.

Video Joint Embedding Predictive Architectures (V-JEPA) learn generalizable off-the-shelf video representation by predicting masked regions in latent space with an exponential moving average (EMA)-updated teacher. While EMA prevents representation collapse, it complicates scalable model selection and couples teacher and student architectures. We revisit masked-latent prediction and show that a frozen teacher suffices. Concretely, we (i) train a target encoder with a simple pixel-reconstruction objective under V-JEPA masking, then (ii) freeze it and train a student to predict the teacher's latents on masked regions. This leads to a two-stage, unregularized scheme that we refer to as SALT (Static-teacher Asymmetric Latent Training). SALT decouples optimization into pixel reconstruction (teacher) and masked latent prediction (student), increasing transparency, efficiency, and scalability while preserving the ability of representation to generalize under frozen evaluation. Empirically, our student models outperform recently proposed V-JEPA 2 encoders under frozen backbone evaluation across diverse benchmarks. They are also more compute-optimal: at matched pretraining FLOPs, our method achieves higher probing accuracy, and its scaling curves dominate V-JEPA's accuracy-FLOPs Pareto frontier. Finally, we find that student quality is remarkably robust to teacher quality: high-performing students emerge even with small, sub-optimal teachers. This points to a compute budget allocation that should overwhelmingly favor the student. These results position SALT as a simple, scalable, and compute-efficient alternative to EMA-based self-distillation for video representation learning.

Test-time compute has emerged as a key paradigm for enhancing LLM reasoning, yet prevailing approaches like Best-of-N and majority voting apply uniform depth across inputs, wasting computation on simple queries while potentially under-thinking complex ones. We present FR-Ponder, a single-graph, backbone-training-free framework that allocates instance-adaptive reasoning compute via latent steering. A less than 1M-param controller observes hidden states and decides to halt or apply a small ponder step by adding a pre-computed steering vector to frozen representations. Our method extracts the latent steering vector associated with deeper reasoning outputs and direct IO from LLM and re-applies it through a tunable scaling factor, allowing the model to adapt its reasoning depth to the complexity of each input. To balance performance and computational cost, we employ Group Relative Policy Optimization (GRPO) as a reward signal to adaptively regulate reasoning depth, achieving task accuracy while mitigating overreasoning. Through curriculum learning and careful reward engineering, FR-Ponder learns calibrated compute allocation correlated with problem difficulty. On GSM8K and MA TH500, FR-Ponder improves the compute-accuracy frontier, delivering lower FLOPs with better matched accuracy and comparing favorably to early-exit baselines, without modifying backbone weights. Analyses visualize interpretable steering directions and show learned compute allocation correlates with problem difficulty.

Robots trained via Reinforcement Learning (RL) or Imitation Learning (IL) often adapt slowly to new tasks, whereas recent Large Language Models (LLMs) and Vision-Language Models (VLMs) promise knowledge-rich planning from minimal data. Deploying LLMs/VLMs for motion planning, however, faces two key obstacles: (i) symbolic plans are rarely grounded in scene geometry and object physics, and (ii) model outputs can vary for identical prompts, undermining execution reliability. We propose ViReSkill, a framework that pairs vision-grounded replanning with a skill memory for accumulation and reuse. When a failure occurs, the replanner generates a new action sequence conditioned on the current scene, tailored to the observed state. On success, the executed plan is stored as a reusable skill and replayed in future encounters without additional calls to LLMs/VLMs. This feedback loop enables autonomous continual learning: each attempt immediately expands the skill set and stabilizes subsequent executions. We evaluate ViReSkill on simulators such as LIBERO and RLBench as well as on a physical robot. Across all settings, it consistently outperforms conventional baselines in task success rate, demonstrating robust sim-to-real generalization.