While Reinforcement Learning with Verifiable Rewards (RLVR) has advanced the reasoning capabilities of Large Vision-Language Models (LVLMs), most existing methods in multimodal reasoning neglect the critical role of visual perception within the RLVR optimization process. In this paper, we undertake a pioneering exploration of multimodal RLVR through the novel perspective of token perception, which measures the visual dependency of each generated token. With a granular analysis of Chain-of-Thought (CoT) processes, we uncover two key insights: first, token perception in a rollout trajectory is sparsely distributed, where only a small fraction of tokens have high visual dependency for visually-grounded reasoning; second, different trajectories exhibit significant divergence in their overall visual dependency. Based on these observations, we propose Visually-Perceptive Policy Optimization (VPPO), a novel policy gradient algorithm that explicitly leverages token perception to refine the learning signal. Specifically, VPPO achieves this through a dual mechanism: it reweights a trajectory's advantage by its overall visual dependency, and focuses policy updates exclusively on perceptually pivotal tokens. On a comprehensive suite of eight perception and reasoning benchmarks, VPPO demonstrates substantial gains over leading open-source RL-tuned models, with its effectiveness consistently validated across 7B and 32B model scales. Our findings not only establish a new token-level perceptual perspective for analyzing multimodal RLVR but also present a novel and effective optimization strategy to significantly enhance the multimodal reasoning capabilities of LVLMs.

Vision-language models have achieved remarkable success in multi-modal representation learning from large-scale pairs of visual scenes and linguistic descriptions. However, they still struggle to simultaneously express two distinct types of semantic structures: the hierarchy within a concept family (e.g., dog $\preceq$ mammal $\preceq$ animal) and the compositionality across different concept families (e.g., "a dog in a car" $\preceq$ dog, car). Recent works have addressed this challenge by employing hyperbolic space, which efficiently captures tree-like hierarchy, yet its suitability for representing compositionality remains unclear. To resolve this dilemma, we propose PHyCLIP, which employs an $\ell_1$-Product metric on a Cartesian product of Hyperbolic factors. With our design, intra-family hierarchies emerge within individual hyperbolic factors, and cross-family composition is captured by the $\ell_1$-product metric, analogous to a Boolean algebra. Experiments on zero-shot classification, retrieval, hierarchical classification, and compositional understanding tasks demonstrate that PHyCLIP outperforms existing single-space approaches and offers more interpretable structures in the embedding space.

Hateful memes have emerged as a particularly challenging form of online abuse, motivating the development of automated detection systems. Most prior approaches rely on direct detection, producing only binary predictions. Such models fail to provide the context and explanations that real-world moderation requires. Recent Explain-then-Detect approaches, using Chain-of-Thought prompting or LMM agents, perform worse than simple SFT baselines, and even advanced post-training methods such as GRPO fail to close the gap. Our analysis identifies two key issues of such systems: important policy-relevant cues such as targets and attack types are not hypothesized by the model as a likely explanation; and the binary reward signal is insufficient to guide reasoning. To address these challenges, we propose ExPO-HM (Explain-then-Detect Policy Optimization for Hateful Memes), inspired by the training and evaluation process of human annotators. ExPO-HM combines SFT warmup, GRPO with curriculum learning, and Conditional Decision Entropy (CDE) as both metric and reward for reasoning quality. Across three hateful meme benchmarks, ExPO-HM achieves state-of-the-art performance on binary detection, fine-grained classification, and reasoning quality, with up to 15\% and 17\% F1 improvement over the GRPO and DPO baselines, respectively. By moving hateful meme detection from simple binary alarms to explanation-driven detection, ExPO-HM provides accurate, interpretable, and actionable moderation support. Code available at https://github.com/JingbiaoMei/ExPO-HM

Central to the widespread use of t-distributed stochastic neighbor embedding (t-SNE) is the conviction that it produces visualizations whose structure roughly matches that of the input. To the contrary, we prove that (1) the strength of the input clustering, and (2) the extremity of outlier points, cannot be reliably inferred from the t-SNE output. We demonstrate the prevalence of these failure modes in practice as well.

Pretrained transformers readily adapt to new sequence modeling tasks via zero-shot prompting, but relational domains still lack architectures that transfer across datasets and tasks. The core challenge is the diversity of relational data, with varying heterogeneous schemas, graph structures and functional dependencies. In this paper, we present the Relational Transformer (RT) architecture, which can be pretrained on diverse relational databases and directly applied to unseen datasets and tasks without task- or dataset-specific fine-tuning, or retrieval of in-context examples. RT (i) incorporates task specification via task table prompting, (ii) tokenizes cells with table/column metadata, (iii) is pretrained via masked token prediction, and (iv) utilizes a novel Relational Attention mechanism over columns, rows, and primary-foreign key links. Pretrained on RelBench datasets spanning tasks such as churn and sales forecasting, RT attains strong zero-shot performance, averaging 93% of fully supervised AUROC on binary classification tasks with a single forward pass of a 22M parameter model, as opposed to 84% for a 27B LLM. Fine-tuning yields state-of-the-art results with high sample efficiency. Our experimental analyses show that RT's zero-shot transfer leverages task context, relational attention patterns and schema semantics. Overall, RT provides a practical path toward foundation models for relational data. Code, models, data: https://github.com/snap-stanford/relational-transformer.

Most existing benchmarks for understanding egocentric vision focus primarily on daytime scenarios, overlooking the low-light conditions that are inevitable in real-world applications. To investigate this gap, we present EgoNight, the first comprehensive benchmark for nighttime egocentric vision, with visual question answering (VQA) as the core task. A key feature of EgoNight is the introduction of day-night aligned videos, which enhance night annotation quality using the daytime data and reveal clear performance gaps between lighting conditions. To achieve this, we collect both synthetic videos rendered by Blender and real-world recordings, ensuring that scenes and actions are visually and temporally aligned. Leveraging these paired videos, we construct EgoNight-VQA, supported by a novel day-augmented night auto-labeling engine and refinement through extensive human verification. Each QA pair is double-checked by annotators for reliability. In total, EgoNight-VQA contains 3658 QA pairs across 90 videos, spanning 12 diverse QA types, with more than 300 hours of human work. Evaluations of state-of-the-art multimodal large language models (MLLMs) reveal substantial performance drops when transferring from day to night, underscoring the challenges of reasoning under low-light conditions. Beyond VQA, EgoNight also introduces two auxiliary tasks, day-night correspondence retrieval and egocentric depth estimation at night, that further explore the boundaries of existing models. We believe EgoNight-VQA provides a strong foundation for advancing application-driven egocentric vision research and for developing models that generalize across illumination domains. The code and data can be found at https://github.com/dehezhang2/EgoNight.

Scaling unsupervised skill discovery algorithms to high-DoF agents remains challenging. As dimensionality increases, the exploration space grows exponentially, while the manifold of meaningful skills remains limited. Therefore, semantic meaningfulness becomes essential to effectively guide exploration in high-dimensional spaces. In this work, we present Reference-Grounded Skill Discovery (RGSD), a novel algorithm that grounds skill discovery in a semantically meaningful latent space using reference data. RGSD first performs contrastive pretraining to embed motions on a unit hypersphere, clustering each reference trajectory into a distinct direction. This grounding enables skill discovery to simultaneously involve both imitation of reference behaviors and the discovery of semantically related diverse behaviors. On a simulated SMPL humanoid with $359$-D observations and $69$-D actions, RGSD successfully imitates skills such as walking, running, punching, and sidestepping, while also discover variations of these behaviors. In downstream locomotion tasks, RGSD leverages the discovered skills to faithfully satisfy user-specified style commands and outperforms imitation-learning baselines, which often fail to maintain the commanded style.

Low-Rank Adaptation (LoRA) has emerged as a dominant method in Parameter-Efficient Fine-Tuning (PEFT) for large language models, which augments the transformer layer with one down-projection $A$ and one up-projection $B$. However, LoRA's reliance on a single down-projection matrix ($A$) creates a representational bottleneck, as this solitary feature extractor is inherently insufficient for capturing the diverse signals required by complex tasks. This motivates our architectural shift to focus on enriching the feature adaptation to improve the downstream task adaptation ability. We propose MASA (Multi-$A$ Shared Adaptation), an architecture that implements a multi-$A$, single-$B$ structure where the multi-$A$ expert ensemble is asymmetrically shared across layers to ensure parameter efficiency. In MASA, these specialized experts capture diverse features, which are then integrated by a single, layer-specific $B$-matrix. The effectiveness and versatility of our method are validated through a comprehensive suite of experiments spanning multi-domain generalization, single-domain specialization, and multi-task reasoning. For example, on the MMLU benchmark, MASA achieves an average accuracy of 59.62%, outperforming the standard LoRA by 1.08 points (a relative improvement of 1.84%) with comparable learnable parameters of 0.52%.

Self-play preference optimization has emerged as a prominent paradigm for aligning large language models (LLMs). It typically involves a language model to generate on-policy responses for prompts and a reward model (RM) to guide the selection of chosen and rejected responses, which can be further trained with direct preference optimization (DPO). However, the role of prompts remains underexplored, despite being a core component in this pipeline. In this work, we investigate how prompts of varying difficulty influence self-play preference optimization. We use the mean reward of sampled responses of a prompt as a proxy for its difficulty. We first find that difficult prompts exhibit substantially inferior self-play optimization performance compared to easy prompts for language models. Moreover, incorporating difficult prompts into training fails to enhance overall performance and, in fact, leads to slight degradation compared to training on easy prompts alone. Third, there is a clear upward trend in optimization performance as prompt difficulty decreases. We also observe that the performance gap between difficult and easy prompts tends to close as the model capacity increases, suggesting that prompt difficulty interacts with the model capacity. Building on these findings, we explore strategies to mitigate the adversary effect of difficult prompts on final performance. We demonstrate that only training on a small portion (30%) of the easiest prompts improves overall self-play performance on AlpacaEval~2 and Arena-Hard. We also report failed attempts and lessons learned.

Although LLMs have demonstrated improved performance by scaling parallel test-time compute, doing so relies on generating reasoning paths that are both diverse and accurate. For challenging problems, the forking tokens that trigger diverse yet correct reasoning modes are typically deep in the sampling tree. Consequently, common strategies to encourage diversity, such as temperature scaling, encounter a worsened trade-off between diversity and accuracy. Motivated by this challenge, we treat parallel reasoning as a set-of-next-token-prediction problem and incorporate a set-based global loss into Supervised Fine-Tuning (SFT) using bipartite matching between global forking tokens and unique reasoning traces. We observe that whereas naive fine-tuning with multiple reasoning traces collapses these unique reasoning modes, our proposed method, Set Supervised Fine-Tuning (SSFT), preserves these modes and produces emergent global forking tokens. Global Forking Policy Optimization (GFPO) leverages these maximally steerable tokens to incentivize complex reasoning, and the resulting models consistently outperform their SFT counterparts with GRPO on both math reasoning and execution-based code generation benchmarks.

Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free settings: 1) purely latent reasoning broadens the search distribution by maintaining multiple implicit paths, which diffuses probability mass, introduces noise, and impedes convergence to a single high-confidence solution, thereby hurting accuracy; and 2) overthinking persists even without explicit text, wasting tokens and degrading efficiency. To address these issues, we introduce SwiReasoning, a training-free framework for LLM reasoning which features two key innovations: 1) SwiReasoning dynamically switches between explicit and latent reasoning, guided by block-wise confidence estimated from entropy trends in next-token distributions, to balance exploration and exploitation and promote timely convergence. 2) By limiting the maximum number of thinking-block switches, SwiReasoning curbs overthinking and improves token efficiency across varying problem difficulties. On widely used mathematics, STEM, coding, and general benchmarks, SwiReasoning consistently improves average accuracy by 1.8%-3.1% across reasoning LLMs of different model families and scales. Furthermore, under constrained budgets, SwiReasoning improves average token efficiency by 57%-79%, with larger gains as budgets tighten.

Conformal prediction offers a powerful framework for building distribution-free prediction intervals for exchangeable data. Existing methods that extend conformal prediction to sequential data rely on fitting a relatively complex model to capture temporal dependencies. However, these methods can fail if the sample size is small and often require expensive retraining when the underlying data distribution changes. To overcome these limitations, we propose Reservoir Conformal Prediction (ResCP), a novel training-free conformal prediction method for time series. Our approach leverages the efficiency and representation learning capabilities of reservoir computing to dynamically reweight conformity scores. In particular, we compute similarity scores among reservoir states and use them to adaptively reweight the observed residuals at each step. With this approach, ResCP enables us to account for local temporal dynamics when modeling the error distribution without compromising computational scalability. We prove that, under reasonable assumptions, ResCP achieves asymptotic conditional coverage, and we empirically demonstrate its effectiveness across diverse forecasting tasks.

Bayesian optimization has emerged as a prominent methodology for optimizing expensive black-box functions by leveraging Gaussian process surrogates, which focus on capturing the global characteristics of the objective function. However, in numerous practical scenarios, the primary objective is not to construct an exhaustive global surrogate, but rather to quickly pinpoint the global optimum. Due to the aleatoric nature of the sequential optimization problem and its dependence on the quality of the surrogate model and the initial design, it is restrictive to assume that all observed samples contribute equally to the discovery of the optimum in this context. In this paper, we introduce Counterfactual Credit Guided Bayesian Optimization (CCGBO), a novel framework that explicitly quantifies the contribution of individual historical observations through counterfactual credit. By incorporating counterfactual credit into the acquisition function, our approach can selectively allocate resources in areas where optimal solutions are most likely to occur. We prove that CCGBO retains sublinear regret. Empirical evaluations on various synthetic and real-world benchmarks demonstrate that CCGBO consistently reduces simple regret and accelerates convergence to the global optimum.

Recent large reasoning models (LRMs) driven by reinforcement learning algorithms (e.g., GRPO) have achieved remarkable performance on challenging reasoning tasks. However, these models suffer from overthinking, generating unnecessarily long and redundant reasoning even for simple questions, which substantially increases computational cost and response latency. While existing methods incorporate length rewards to GRPO to promote concise reasoning, they incur significant performance degradation. We identify the root cause: when rewards for correct but long rollouts are penalized, GRPO's group-relative advantage function can assign them negative advantages, actively discouraging valid reasoning. To overcome this, we propose Decoupled Reward Policy Optimization (DRPO), a novel framework that decouples the length-based learning signal of correct rollouts from incorrect ones. DRPO ensures that reward signals for correct rollouts are normalized solely within the positive group, shielding them from interference by negative samples. The DRPO's objective is grounded in integrating an optimized positive data distribution, which maximizes length-based rewards under a KL regularization, into a discriminative objective. We derive a closed-form solution for this distribution, enabling efficient computation of the objective and its gradients using only on-policy data and importance weighting. Of independent interest, this formulation is general and can incorporate other preference rewards of positive data beyond length. Experiments on mathematical reasoning tasks demonstrate DRPO's significant superiority over six efficient reasoning baselines. Notably, with a 1.5B model, our method achieves 77\% length reduction with only 1.1\% performance loss on simple questions like GSM8k dataset, while the follow-up baseline sacrifices 4.3\% for 68\% length reduction.

Test-time scaling improves the reasoning capabilities of large language models (LLMs) by allocating extra compute to generate longer Chains-of-Thoughts (CoTs). This enables models to tackle more complex problem by breaking them down into additional steps, backtracking, and correcting mistakes. Despite its strong performance--demonstrated by OpenAI's o1 and DeepSeek R1, the conditions in the training data under which long CoTs emerge, and when such long CoTs improve the performance, remain unclear. In this paper, we study the performance of test-time scaling for transformers trained on an in-context weight prediction task for linear regression. Our analysis provides a theoretical explanation for several intriguing observations: First, at any fixed test error, increasing test-time compute allows us to reduce the number of in-context examples (context length) in training prompts. Second, if the skills required to solve a downstream task are not sufficiently present in the training data, increasing test-time compute can harm performance. Finally, we characterize task hardness via the smallest eigenvalue of its feature covariance matrix and show that training on a diverse, relevant, and hard set of tasks results in best performance for test-time scaling. We confirm our findings with experiments on large, nonlinear transformer architectures.

Large Language Models (LLMs) as autonomous agents are increasingly tasked with solving complex, long-horizon problems. Aligning these agents via preference-based offline methods like Direct Preference Optimization (DPO) is a promising direction, yet it faces a critical granularity mismatch. Trajectory-level DPO provides a signal that is too coarse for precise credit assignment, while step-level DPO is often too myopic to capture the value of multi-step behaviors. To resolve this challenge, we introduce Hierarchical Preference Learning (HPL), a hierarchical framework that optimizes LLM agents by leveraging preference signals at multiple, synergistic granularities. While HPL incorporates trajectory- and step-level DPO for global and local policy stability, its core innovation lies in group-level preference optimization guided by a dual-layer curriculum. Our approach first decomposes expert trajectories into semantically coherent action groups and then generates contrasting suboptimal groups to enable preference learning at a fine-grained, sub-task level. Then, instead of treating all preference pairs equally, HPL introduces a curriculum scheduler that organizes the learning process from simple to complex. This curriculum is structured along two axes: the group length, representing sub-task complexity, and the sample difficulty, defined by the reward gap between preferred and dispreferred action groups. Experiments on three challenging agent benchmarks show that HPL outperforms existing state-of-the-art methods. Our analyses demonstrate that the hierarchical DPO loss effectively integrates preference signals across multiple granularities, while the dual-layer curriculum is crucial for enabling the agent to solve a wide range of tasks, from simple behaviors to complex multi-step sequences.

Drag-based image editing has long suffered from distortions in the target region, largely because the priors of earlier base models, Stable Diffusion, are insufficient to project optimized latents back onto the natural image manifold. With the shift from UNet-based DDPMs to more scalable DiT with flow matching (e.g., SD3.5, FLUX), generative priors have become significantly stronger, enabling advances across diverse editing tasks. However, drag-based editing has yet to benefit from these stronger priors. This work proposes the first framework to effectively harness FLUX's rich prior for drag-based editing, dubbed DragFlow, achieving substantial gains over baselines. We first show that directly applying point-based drag editing to DiTs performs poorly: unlike the highly compressed features of UNets, DiT features are insufficiently structured to provide reliable guidance for point-wise motion supervision. To overcome this limitation, DragFlow introduces a region-based editing paradigm, where affine transformations enable richer and more consistent feature supervision. Additionally, we integrate pretrained open-domain personalization adapters (e.g., IP-Adapter) to enhance subject consistency, while preserving background fidelity through gradient mask-based hard constraints. Multimodal large language models (MLLMs) are further employed to resolve task ambiguities. For evaluation, we curate a novel Region-based Dragging benchmark (ReD Bench) featuring region-level dragging instructions. Extensive experiments on DragBench-DR and ReD Bench show that DragFlow surpasses both point-based and region-based baselines, setting a new state-of-the-art in drag-based image editing. Code and dataset are available at https://github.com/Edennnnnnnnnn/DragFlow.

The dominant paradigm for training large reasoning models starts with pre-training using next-token prediction loss on vast amounts of data. Reinforcement learning, while powerful in scaling reasoning, is introduced only as the very last phase of post-training, preceded by supervised fine-tuning. While dominant, is this an optimal way of training? In this paper, we present RLP, an information-driven reinforcement pretraining objective, that brings the core spirit of reinforcement learning -- exploration -- to the last phase of pretraining. The key idea is to treat chain-of-thought as an exploratory action, with rewards computed based on the information gain it provides for predicting future tokens. This training objective essentially encourages the model to think for itself before predicting what comes next, thus teaching an independent thinking behavior earlier in the pretraining. More concretely, the reward signal measures the increase in log-likelihood of the next token when conditioning on both context and a sampled reasoning chain, compared to conditioning on context alone. This approach yields a verifier-free dense reward signal, allowing for efficient training for the full document stream during pretraining. Specifically, RLP reframes reinforcement learning for reasoning as a pretraining objective on ordinary text, bridging the gap between next-token prediction and the emergence of useful chain-of-thought reasoning. Pretraining with RLP on Qwen3-1.7B-Base lifts the overall average across an eight-benchmark math-and-science suite by 19%. With identical post-training, the gains compound, with the largest improvements on reasoning-heavy tasks such as AIME25 and MMLU-Pro. Applying RLP to the Nemotron-Nano-12B-v2 increases the overall average from 42.81% to 61.32% and raises the average on scientific reasoning by 23%, demonstrating scalability across architectures and model sizes.

The training paradigm for large language models (LLMs) is moving from static datasets to experience-based learning, where agents acquire skills via interacting with complex environments. To facilitate this transition we introduce GEM (General Experience Maker), an open-source environment simulator designed for the age of LLMs. Analogous to OpenAI-Gym for traditional reinforcement learning (RL), GEM provides a standardized framework for the environment-agent interface, including asynchronous vectorized execution for high throughput, and flexible wrappers for easy extensibility. GEM also features a diverse suite of environments, robust integrated tools, and single-file example scripts demonstrating using GEM with five popular RL training frameworks. Along with this, we also provide a set of baselines across 24 environments using REINFORCE with Return Batch Normalization (ReBN), which -- unlike GRPO -- is compatible with the full RL setting of dense per-turn rewards and offers better credit assignment. We further conduct apple-to-apple benchmarking of PPO, GRPO and REINFORCE in both single- and multi-turn settings using GEM to shed light on the algorithmic designs. Lastly, GEM also functions as a convenient evaluation toolkit besides a training environment. We hope this framework can help accelerate future agentic LLM research.

Training algorithms in deep learning usually treat a mini-batch of samples as a single object; they average gradients over the mini-batch, and then process the average in various ways. Computing other statistics beyond the average may have been seen as prohibitively resource intensive in automatic differentiation (AD) frameworks. We show that this is not the case. Generally, gradient statistics can be implemented through a surgery of the AD graph, which, in some cases, incur almost no computational and memory overheads compared to the mini-batch gradient computation. Additionally, we show that in certain classes of models, including transformers, JAX's vectorization transformation offers a viable implementation for prototyping and experimentation. We then revise our understanding of two nonlinear operations in optimization through the lens of per-example gradient transformations. We first study signSGD and show that the optimal placement of the sign operation in the gradient processing chain is crucial to success and can be predicted with a simple signal-to-noise ratio argument. Next we study per-example variations of the Adam preconditioner, and show that optimization is best served when the preconditioner is dominated by the mean rather than the variance of the gradient distribution - in contrast to conventional wisdom. Overall we demonstrate that per-example gradient information enables new analyses and possibilities for algorithm design.

Diffusion-based large language models (dLLMs) are gaining attention for their inherent capacity for parallel decoding, offering a compelling alternative to autoregressive LLMs. Among various decoding strategies, block-wise semi-autoregressive (semi-AR) approaches are widely adopted due to their support for KV caching and their favorable accuracy-speed trade-off. However, this paper identifies two fundamental limitations in the conventional semi-AR decoding approach that applies a fixed block size: i) late decoding overhead, where the unmasking of high-confidence tokens outside the current block is unnecessarily delayed, and ii) premature decoding error, where low-confidence tokens inside the current block are committed too early, leading to incorrect tokens. This paper presents the first systematic investigation challenging the fixed block size setting in semi-AR decoding. Through a statistical analysis of confidence dynamics during the denoising process, we identify a volatility band (VB) region during dLLM decoding, which encodes local semantic structure and can be used to guide adaptive block sizing. Leveraging these insights, we introduce AdaBlock-dLLM, a training-free, plug-and-play scheduler that adaptively aligns block boundaries with semantic steps by adjusting block size during runtime. Extensive experiments across diverse benchmarks show that AdaBlock-dLLM achieves up to 5.3% accuracy improvement under the same throughput budget. Beyond inference-time optimization, we hope our semantics-aware adaptive scheduling approach and confidence-based analysis will inspire future training strategies for dLLMs. Our code is available at https://github.com/lgxi24/AdaBlock-dLLM.

Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM's optimized inference engine, EasySteer achieves 10.8-22.3$\times$ speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.

Effective LLM training depends on predictable scaling of key quantities -- such as final loss and optimal hyperparameters -- with model and dataset size. Qiu et al. (2025) recently showed that this predictability can extend beyond scalars: whole training loss curves can *collapse* onto a universal trajectory after a simple normalization. What remains unclear is whether this phenomenon persists for LLM families trained under *practical scaling recipes*, where width, depth, learning rate, batch size, and weight decay are scaled jointly. We show that it does: loss curves collapse across scales precisely when optimization hyperparameters are set optimally for the given data budget, in accordance with recent empirical scaling laws. Collapse therefore emerges as a signature of compute-efficient training. We demonstrate two applications at scale: (1) deviation-from-collapse provides a sensitive, early diagnostic of training pathologies, and (2) predictability of collapsed curves enables early stopping in large-scale hyperparameter tuning. Finally, we train a competitive LLM family, *Celerity*, using these insights, establishing collapse as an effective tool for developing efficient LLMs.

Knowledge editing aims to efficiently correct factual errors in language models. Widely used locate-then-edit methods update an MLP layer by adjusting its weights to change the mapping between the layer's input vector (key) and output vector (value), thereby editing the model's knowledge. As this update is driven by key and value vectors, obtaining these vectors without careful constraints causes significant model perturbations beyond the targeted edit, a common issue in many prior knowledge editing methods. To address this, we propose Subspace Knowledge Edit (SUIT), which computes key and value vectors only within the subspace of critical features relevant to the edit. Our empirical results on LLaMA3, GPT-J, and Qwen2.5 models show that SUIT dramatically improves knowledge preservation over strong baselines while maintaining high editing performance. These results support the claim that SUIT successfully identifies the critical subspace for the edit. Beyond quantitative gains, our analyses show that SUIT reduces unintended perturbations in hidden states while confining updates to directions that are more effective for editing. Taken together, these findings establish edit-critical subspace identification as a key principle for reliable, low-perturbation knowledge editing. Our code is available at https://github.com/holi-lab/SUIT.

Recent developments in Multimodal Large Language Models (MLLMs) have significantly improved Vision-Language (VL) reasoning in 2D domains. However, extending these capabilities to 3D scene understanding remains a major challenge. Existing 3D Multimodal Large Language Models (3D-MLLMs) often depend on 3D data inputs, which limits scalability and generalization. To address this limitation, we propose Vid-LLM, a video-based 3D-MLLM that directly processes video inputs without requiring external 3D data, making it practical for real-world deployment. In our method, the geometric prior are directly used to improve the performance of the sceen perception. To integrate the geometric cues into the MLLM compactly, we design a Cross-Task Adapter (CTA) module to align the 3D geometric priors with the vision-language representations. To ensure geometric consistency and integrity, we introduce a Metric Depth Model that recovers real-scale geometry from the reconstruction outputs. Finally, the model is fine-tuned with a two-stage distillation optimization strategy, realizing fast convergence and stabilizes training. Extensive experiments across diverse benchmarks verified the effectiveness of our method on 3D Question Answering, 3D Dense Captioning and 3D Visual Grounding tasks, demonstrating the superior multi-task capabilities.

Unified Multimodal Models (UMMs) built on shared autoregressive (AR) transformers are attractive for their architectural simplicity. However, we identify a critical limitation: when trained on multimodal inputs, modality-shared transformers suffer from severe gradient conflicts between vision and text, particularly in shallow and deep layers. We trace this issue to the fundamentally different low-level statistical properties of images and text, while noting that conflicts diminish in middle layers where representations become more abstract and semantically aligned. To overcome this challenge, we propose Uni-X, a two-end-separated, middle-shared architecture. Uni-X dedicates its initial and final layers to modality-specific processing, while maintaining shared parameters in the middle layers for high-level semantic fusion. This X-shaped design not only eliminates gradient conflicts at both ends but also further alleviates residual conflicts in the shared layers. Extensive experiments validate the effectiveness of Uni-X. Under identical training conditions, Uni-X achieves superior training efficiency compared to strong baselines. When scaled to 3B parameters with larger training data, Uni-X matches or surpasses 7B AR-based UMMs, achieving a GenEval score of 82 for image generation alongside strong performance in text and vision understanding tasks. These results establish Uni-X as a parameter-efficient and scalable foundation for future unified multimodal modeling. Our code is available at https://github.com/CURRENTF/Uni-X

We introduce $\textbf{SimuHome}$, a high-fidelity smart home simulator and a benchmark of 600 episodes for LLM-based smart home agents. Existing smart home benchmarks treat the home as a static system, neither simulating how device operations affect environmental variables over time nor supporting workflow scheduling of device commands. SimuHome is grounded in the Matter protocol, the industry standard that defines how real smart home devices communicate and operate. Agents interact with devices through SimuHome's APIs and observe how their actions continuously affect environmental variables such as temperature and humidity. Our benchmark covers state inquiry, implicit user intent inference, explicit device control, and workflow scheduling, each with both feasible and infeasible requests. For workflow scheduling, the simulator accelerates time so that scheduled workflows can be evaluated immediately. An evaluation of 18 agents reveals that workflow scheduling is the hardest category, with failures persisting across alternative agent frameworks and fine-tuning. These findings suggest that SimuHome's time-accelerated simulation could serve as an environment for agents to pre-validate their actions before committing them to the real world.

Prompt optimization and fine-tuning are two major approaches to improve the performance of Large Language Models (LLMs). They enhance the capabilities of LLMs from complementary perspectives: the former through explicit natural language, and the latter through implicit parameter updates. However, prior work has typically studied them in isolation, leaving their synergistic potential largely underexplored. To bridge this gap, in this paper, we introduce MetaTuner, a novel framework that jointly integrates prompt optimization and fine-tuning for LLM training. Specifically, we introduce two neural networks to generate prompts and parameters, respectively, while allowing them to share a common bottom encoding layer to enable knowledge sharing. By the guidance of the final supervised signals, our framework is optimized to discover the optimal combinations between the prompts and parameters. Given that prompt learning involves discrete optimization while fine-tuning operates in a continuous parameter space, we design a supervised regularization loss to train our framework effectively. Extensive experiments across diverse benchmarks show that our method consistently outperforms the baselines.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities. The code is available: https://github.com/ZJUWYH/FARL.

Dynamic manipulation is a key capability for advancing robot performance, enabling skills such as tossing. While recent learning-based approaches have pushed the field forward, most methods still rely on manually designed action parameterizations, limiting their ability to produce the highly coordinated motions required in complex tasks. Motion planning can generate feasible trajectories, but the dynamics gap-stemming from control inaccuracies, contact uncertainties, and aerodynamic effects-often causes large deviations between planned and executed trajectories. In this work, we propose Dynamics-Aware Motion Manifold Primitives (DA-MMP), a motion generation framework for goal-conditioned dynamic manipulation, and instantiate it on a challenging real-world ring-tossing task. Our approach extends motion manifold primitives to variable-length trajectories through a compact parameterization and learns a high-quality manifold from a large-scale dataset of planned motions. Building on this manifold, a conditional flow matching model is trained in the latent space with a small set of real-world trials, enabling the generation of throwing trajectories that account for execution dynamics. Experiments show that our method can generate coordinated and smooth motion trajectories for the ring-tossing task. In real-world evaluations, it achieves high success rates and even surpasses the performance of trained human experts. Moreover, it generalizes to novel targets beyond the training range, indicating that it successfully learns the underlying trajectory-dynamics mapping.