Human decision-making often involves constrained optimization. As LLM agents are deployed to assist with real-world tasks like travel planning, shopping, and scheduling, they must mirror this capability. We introduce COMPASS, a benchmark that evaluates whether LLM agents can perform constrained optimization in realistic travel planning settings. To success in these tasks, agents must engage in multi-turn conversations with user to gather task information as well as use tools to gather information from the database. Then agents must propose a solution that not only satisfies hard constraints but also optimizes user's utility objective. Evaluating state-of-the-art models, we reveal a significant feasible-optimal gap: while models achieve 70-90% feasibility (constraint satisfaction), they reach only 20-60% optimality (utility optimization). Our analysis shows that tool use is not the bottleneck. Instead, the core limitation is insufficient exploration of the search space, with success strongly correlating with information gathered. Coding agents show a promising approach to mitigate this gap. Together, COMPASS provides a testbed for developing LLM agents that can truly mirror human decision-making by both satisfying constraints and optimizing objectives.

The Uniform Information Density (UID) hypothesis proposes that effective communication is achieved by maintaining a stable flow of information. In this work, we revisit this principle in the context of Large Language Model (LLM) reasoning, asking whether step-level uniformity reflects reasoning quality. To this end, we introduce a novel framework to quantify uniformity of information flow at both local and global levels, using an entropy-based stepwise density metric. Across experiments on seven reasoning benchmarks, we see a counter-intuitive pattern: while high-quality reasoning exhibit smooth step-by-step transitions local uniformity and structured, non-uniform information flow at the trajectory level global non-uniformity. The results demonstrate that these uniformities outperform alternative internal signals as predictors of reasoning quality, and such divergence with human communication is not a model deficiency, but a byproduct of distinct objectives between human communication and LLM reasoning.

People often encounter role conflicts -- social dilemmas where the expectations of multiple roles clash and cannot be simultaneously fulfilled. As large language models (LLMs) increasingly navigate these social dynamics, a critical research question emerges. When faced with such dilemmas, do LLMs prioritize dynamic contextual cues or the learned preferences? To address this, we introduce RoleConflictBench, a novel benchmark designed to measure the contextual sensitivity of LLMs in role conflict scenarios. To enable objective evaluation within this subjective domain, we employ situational urgency as a constraint for decision-making. We construct the dataset through a three-stage pipeline that generates over 13,000 realistic scenarios across 65 roles in five social domains by systematically varying the urgency of competing situations. This controlled setup enables us to quantitatively measure contextual sensitivity, determining whether model decisions align with the situational contexts or are overridden by the learned role preferences. Our analysis of 10 LLMs reveals that models substantially deviate from this objective baseline. Instead of responding to dynamic contextual cues, their decisions are predominantly governed by the preferences toward specific social roles.

The expanding long-context capabilities of large language models are constrained by a significant memory bottleneck: the key-value (KV) cache required for autoregressive generation. This bottleneck is substantial; for instance, a Llama-3.1-8B model processing a 32K-token prompt at a batch size of 4 requires approximately 16GB for its KV cache, a size exceeding the model's weights. While KV-cache compression via low-rank projection is a promising direction, existing methods rely on a static, offline-learned subspace that performs poorly under data distribution shifts. To overcome these limitations, we introduce OjaKV, a novel framework that integrates a strategic hybrid storage policy with online subspace adaptation. First, OjaKV recognizes that not all tokens are equally important for compression; it preserves the crucial first and most recent tokens in full-rank, maintaining high-fidelity anchors for attention. Second, for the vast majority of intermediate tokens, it applies low-rank compression by incrementally adapting the projection basis using Oja's algorithm for online principal component analysis. This adaptation involves a comprehensive update during prompt prefilling and lightweight periodic updates during decoding, ensuring the subspace remains aligned with the evolving context. Crucially, our framework is fully compatible with modern attention modules like FlashAttention. Experiments demonstrate that OjaKV maintains or even improves zero-shot accuracy at high compression ratios. In particular, OjaKV achieves its strongest gains on very long-context benchmarks that require complex reasoning, highlighting the importance of online subspace adaptation in dynamically tracking context shifts. These results establish our hybrid framework as a practical, plug-and-play solution for memory-efficient long-context inference without requiring model fine-tuning.

The de novo generation of drug-like molecules capable of inducing desirable phenotypic changes is receiving increasing attention. However, previous methods predominantly rely on expression profiles to guide molecule generation, but overlook the perturbative effect of the molecules on cellular contexts. To overcome this limitation, we propose SmilesGEN, a novel generative model based on variational autoencoder (VAE) architecture to generate molecules with potential therapeutic effects. SmilesGEN integrates a pre-trained drug VAE (SmilesNet) with an expression profile VAE (ProfileNet), jointly modeling the interplay between drug perturbations and transcriptional responses in a common latent space. Specifically, ProfileNet is imposed to reconstruct pre-treatment expression profiles when eliminating drug-induced perturbations in the latent space, while SmilesNet is informed by desired expression profiles to generate drug-like molecules. Our empirical experiments demonstrate that SmilesGEN outperforms current state-of-the-art models in generating molecules with higher degree of validity, uniqueness, novelty, as well as higher Tanimoto similarity to known ligands targeting the relevant proteins. Moreover, we evaluate SmilesGEN for scaffold-based molecule optimization and generation of therapeutic agents, and confirmed its superior performance in generating molecules with higher similarity to approved drugs. SmilesGEN establishes a robust framework that leverages gene signatures to generate drug-like molecules that hold promising potential to induce desirable cellular phenotypic changes. The source code and datasets are available at: https://github.com/hliulab/SmilesGEN.

Reinforcement Learning with Human Feedback (RLHF) has become crucial for aligning Large Language Models (LLMs) with human intent. However, existing offline RLHF approaches suffer from overoptimization, where language models degrade by overfitting inaccuracies and drifting from preferred behaviors observed during training. Distributionally robust optimization (DRO) is a natural solution, but existing DRO-DPO methods are sample-inefficient, ignore heterogeneous preferences, and lean on brittle heuristics. We introduce \emph{DRO-REBEL}, a family of robust online REBEL updates built on type-$p$ Wasserstein, Kullback-Leibler (KL), and $χ^2$ ambiguity sets. Strong duality reduces each update to a relative-reward regression, retaining REBEL's scalability without PPO-style clipping or value networks. Under linear rewards, log-linear policies, and a standard coverage condition, we prove $\widetilde{O}(\sqrt{d/n})$ bounds on squared parameter error, with sharper constants than prior DRO-DPO analyses, and give the first parametric $\widetilde{O}(d/n)$ rate for DRO-based alignment under preference shift, matching non-robust RLHF in benign regimes. Each divergence yields a tractable SGD-based algorithm: gradient regularization for Wasserstein, importance weighting for KL, and a 1-D dual solve for $χ^2$. On Emotion Alignment, the ArmoRM multi-objective benchmark, and HH-Alignment, DRO-REBEL outperforms prior robust and non-robust baselines across unseen preference mixtures, model sizes, and dataset scales.

Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remains challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We propose a novel framework that combines Large Language Models (LLMs) with a Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components, which are processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances are integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggest a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP. The dataset and implementation details are publicly accessible.

Financial bubbles often arrive without much warning, but create long-lasting economic effects. For example, during the dot-com bubble, innovative technologies created market disruptions through excitement for a promised bright future. Such technologies originated from research where scientists had developed them for years prior to their entry into the markets. That raises a question on the possibility of analyzing scientific publishing data (e.g. citation networks) leading up to a bubble for signals that may forecast the rise and fall of similar future bubbles. To that end, we utilized temporal SNAs to detect possible relationships between the publication citation networks of scientists and financial market data during two modern eras of rapidly shifting technology: 1) dot-com era from 1994 to 2001 and 2) AI era from 2017 to 2024. Results showed that the patterns from the dot-com era (which did end in a bubble) did not definitively predict the rise and fall of an AI bubble. While yearly citation networks reflected possible changes in publishing behavior of scientists between the two eras, there was a subset of AI era scientists whose publication influence patterns mirrored those during the dot-com era. Upon further analysis using multiple analysis techniques (LSTM, KNN, AR X/GARCH), the data seems to suggest two possibilities for the AI era: unprecedented form of financial bubble unseen or that no bubble exists. In conclusion, our findings imply that the patterns present in the dot-com era do not effectively translate in such a manner to apply them to the AI market.

Reinforcement learning with verifiable rewards has significantly advanced reasoning in large language models (LLMs), but such signals remain coarse, offering only binary correctness feedback. This limitation often results in inefficiencies, including overly verbose reasoning and high computational cost, while existing solutions often compromise accuracy. To address this, we introduce self-aligned reward (SAR), a self-guided signal that complements verifiable rewards to encourage both reasoning accuracy and efficiency. SAR is defined as the relative perplexity difference between an answer conditioned on the query and the standalone answer, thereby favoring responses that are concise and query-specific. Quantitative analysis reveals that SAR reliably distinguishes answer quality: concise, correct answers score higher than redundant ones, and partially correct answers score higher than entirely incorrect ones. Evaluation on 4 models across 7 benchmarks shows that integrating SAR with prevalent RL algorithms like PPO and GRPO improves accuracy by 4%, while reducing inference cost by 30%. Further analysis demonstrates that SAR achieves a Pareto-optimal trade-off between correctness and efficiency compared to reward signals based on length or self-confidence. We also show that SAR shortens responses while preserving advanced reasoning behaviors, demonstrating its ability to suppress unnecessary elaboration without losing critical reasoning. These results highlight the promise of self-aligned reward as a fine-grained complement to verifiable rewards, paving the way for more efficient and effective LLM training.

Vision-Language-Action (VLA) models in autonomous driving systems have recently demonstrated transformative potential by integrating multimodal perception with decision-making capabilities. However, the interpretability and coherence of the decision process and the plausibility of action sequences remain largely underexplored. To address these issues, we propose AutoDrive-R$^2$, a novel VLA framework that enhances both reasoning and self-reflection capabilities of autonomous driving systems through chain-of-thought (CoT) processing and reinforcement learning (RL). Specifically, we first propose an innovative CoT dataset named nuScenesR$^2$-6K for supervised fine-tuning, which effectively builds cognitive bridges between input information and output trajectories through a four-step logical chain with self-reflection for validation. Moreover, to maximize both reasoning and self-reflection during the RL stage, we further employ the Group Relative Policy Optimization (GRPO) algorithm within a physics-grounded reward framework that incorporates spatial alignment, vehicle dynamic, and temporal smoothness criteria to ensure reliable and realistic trajectory planning. Extensive evaluation results across both nuScenes and Waymo datasets demonstrates the state-of-the-art performance and robust generalization capacity of our proposed method.

Misleading visualizations are a potent driver of misinformation on social media and the web. By violating chart design principles, they distort data and lead readers to draw inaccurate conclusions. Prior work has shown that both humans and multimodal large language models (MLLMs) are frequently deceived by such visualizations. Automatically detecting misleading visualizations and identifying the specific design rules they violate could help protect readers and reduce the spread of misinformation. However, the training and evaluation of AI models has been limited by the absence of large, diverse, and openly available datasets. In this work, we introduce Misviz, a benchmark of 2,604 real-world visualizations annotated with 12 types of misleaders. To support model training, we also create Misviz-synth, a synthetic dataset of 57,665 visualizations generated using Matplotlib and based on real-world data tables. We perform a comprehensive evaluation on both datasets using state-of-the-art MLLMs, rule-based systems, and image-axis classifiers. Our results reveal that the task remains highly challenging. We release Misviz, Misviz-synth, and the accompanying code.

Achieving both safety guarantees and real-time performance in cooperative vehicle coordination remains a fundamental challenge, particularly in dynamic and uncertain environments. Existing methods often suffer from insufficient uncertainty treatment in safety modeling, which intertwines with the heavy computational burden under complex multi-vehicle coupling. This paper presents a novel coordination framework that resolves this challenge through three key innovations: 1) direct control of vehicles' trajectory distributions during coordination, formulated as a robust cooperative planning problem with adaptive enhanced safety constraints, ensuring a specified level of safety regarding the uncertainty of the interactive trajectory, 2) a fully parallel ADMM-based distributed trajectory negotiation (ADMM-DTN) algorithm that efficiently solves the optimization problem while allowing configurable negotiation rounds to balance solution quality and computational resources, and 3) an interactive attention mechanism that selectively focuses on critical interactive participants to further enhance computational efficiency. Simulation results demonstrate that our framework achieves significant advantages in safety (reducing collision rates by up to 40.79\% in various scenarios) and real-time performance compared to representative benchmarks, while maintaining strong scalability with increasing vehicle numbers. The proposed interactive attention mechanism further reduces the computational demand by 15.4\%. Real-world experiments further validate robustness and real-time feasibility with unexpected dynamic obstacles, demonstrating reliable coordination in complex traffic scenes. The experiment demo could be found at https://youtu.be/4PZwBnCsb6Q.

Unmanned Aerial Vehicles (UAVs) have become increasingly important in disaster emergency response by facilitating aerial video analysis. Due to the limited computational resources available on UAVs, large models cannot be run efficiently for on-board analysis. To overcome this challenge, we propose a lightweight and efficient two-stage framework for wildfire monitoring and fire source detection on UAV platforms. Specifically, in Stage 1, we utilize a policy network to identify and discard redundant video clips, thereby reducing computational costs. We also introduce a station point mechanism that incorporates future frame information within the sequential policy network to improve prediction accuracy. This mechanism allows Stage 1 to operate in a near-real-time manner. In Stage 2, for frames classified as containing fire, we apply an improved YOLOv8 model to accurately localize the fire source in real-time on selected frames. We evaluate Stage 1 using the FLAME and HMDB51 datasets, and Stage 2 using the Fire & Smoke Detection Dataset. Experimental results show that our method significantly reduces computational costs while maintaining classification accuracy in Stage 1, and achieves high detection accuracy with real-time inference in Stage 2.

Unsupervised video Object-Centric Learning (OCL) is promising as it enables object-level scene representation and understanding as we humans do. Mainstream video OCL methods adopt a recurrent architecture: An aggregator aggregates current video frame into object features, termed slots, under some queries; A transitioner transits current slots to queries for the next frame. This is an effective architecture but all existing implementations both (\textit{i1}) neglect to incorporate next frame features, the most informative source for query prediction, and (\textit{i2}) fail to learn transition dynamics, the knowledge essential for query prediction. To address these issues, we propose Random Slot-Feature pair for learning Query prediction (RandSF.Q): (\textit{t1}) We design a new transitioner to incorporate both slots and features, which provides more information for query prediction; (\textit{t2}) We train the transitioner to predict queries from slot-feature pairs randomly sampled from available recurrences, which drives it to learn transition dynamics. Experiments on scene representation demonstrate that our method surpass existing video OCL methods significantly, e.g., up to 10 points on object discovery, setting new state-of-the-art. Such superiority also benefits downstream tasks like scene understanding. Source Code, Model Checkpoints, Training Logs: https://github.com/Genera1Z/RandSF.Q

Multi-Armed Bandit (MAB) algorithms are widely used in recommender systems that require continuous, incremental learning. A core aspect of MABs is the exploration-exploitation trade-off: choosing between exploiting items likely to be enjoyed and exploring new ones to gather information. In contextual linear bandits, this trade-off is particularly central, as many variants share the same linear regression backbone and differ primarily in their exploration strategies. Despite its prevalent use, offline evaluation of MABs is increasingly recognized for its limitations in reliably assessing exploration behavior. This study conducts an extensive offline empirical comparison of several linear MABs. Strikingly, across over 90% of various datasets, a greedy linear model, with no type of exploration, consistently achieves top-tier performance, often outperforming or matching its exploratory counterparts. This observation is further corroborated by hyperparameter optimization, which consistently favors configurations that minimize exploration, suggesting that pure exploitation is the dominant strategy within these evaluation settings. Our results expose significant inadequacies in offline evaluation protocols for bandits, particularly concerning their capacity to reflect true exploratory efficacy. Consequently, this research underscores the urgent necessity for developing more robust assessment methodologies, guiding future investigations into alternative evaluation frameworks for interactive learning in recommender systems.

Improving the reliability of large language models (LLMs) is critical for deploying them in real-world scenarios. In this paper, we propose \textbf{Deliberative Searcher}, the first framework to integrate certainty calibration with retrieval-based search for open-domain question answering. The agent performs multi-step reflection and verification over Wikipedia data and is trained with a reinforcement learning algorithm that optimizes for accuracy under a soft reliability constraint. Empirical results show that proposed method improves alignment between model confidence and correctness, leading to more trustworthy outputs. This paper will be continuously updated.

Successful human-agent teaming relies on an agent being able to understand instructions given by a (human) principal. In many cases, an instruction may be incomplete or ambiguous. In such cases, the agent must infer the unspoken intentions from their shared context, that is, it must exercise the principal's Theory of Mind (ToM) and infer the mental states of its principal. We consider the prospects of effective human-agent collaboration using large language models (LLMs). To assess ToM in a dynamic, goal-oriented, and collaborative environment, we introduce a novel task, Instruction Inference, in which an agent assists a principal in reaching a goal by interpreting incomplete or ambiguous instructions. We present Tomcat, an LLM-based agent, designed to exhibit ToM reasoning in interpreting and responding to the principal's instructions. We implemented two variants of Tomcat. One, dubbed Fs-CoT (Fs for few-shot, CoT for chain-of-thought), is based on a small number of examples demonstrating the requisite structured reasoning. One, dubbed CP (commonsense prompt), relies on commonsense knowledge and information about the problem. We realized both variants of Tomcat on three leading LLMs, namely, GPT-4o, DeepSeek-R1, and Gemma-3-27B. To evaluate the effectiveness of Tomcat, we conducted a study with 52 human participants in which we provided participants with the same information as the CP variant. We computed intent accuracy, action optimality, and planning optimality to measure the ToM capabilities of Tomcat and our study participants. We found that Tomcat with Fs-CoT, particularly with GPT-4o and DeepSeek-R1, achieves performance comparable to the human participants, underscoring its ToM potential for human-agent collaboration.

Reasoning in humans is prone to biases due to underlying motivations like identity protection, that undermine rational decision-making and judgment. This \textit{motivated reasoning} at a collective level can be detrimental to society when debating critical issues such as human-driven climate change or vaccine safety, and can further aggravate political polarization. Prior studies have reported that large language models (LLMs) are also susceptible to human-like cognitive biases, however, the extent to which LLMs selectively reason toward identity-congruent conclusions remains largely unexplored. Here, we investigate whether assigning 8 personas across 4 political and socio-demographic attributes induces motivated reasoning in LLMs. Testing 8 LLMs (open source and proprietary) across two reasoning tasks from human-subject studies -- veracity discernment of misinformation headlines and evaluation of numeric scientific evidence -- we find that persona-assigned LLMs have up to 9% reduced veracity discernment relative to models without personas. Political personas specifically are up to 90% more likely to correctly evaluate scientific evidence on gun control when the ground truth is congruent with their induced political identity. Prompt-based debiasing methods are largely ineffective at mitigating these effects. Taken together, our empirical findings are the first to suggest that persona-assigned LLMs exhibit human-like motivated reasoning that is hard to mitigate through conventional debiasing prompts -- raising concerns of exacerbating identity-congruent reasoning in both LLMs and humans.

Large-scale pre-trained video diffusion models have exhibited remarkable capabilities in diverse video generation. However, existing solutions face several challenges in generating long videos with rich human-scene interactions (HSI), including unrealistic dynamics and affordance, lack of subject identity preservation, and the need for expensive training. To this end, we propose GenHSI, a training-free method for controllable generation of long HSI videos with 3D awareness. Taking inspiration from movie animation, we subdivide the video synthesis into three stages: (1) script writing, (2) pre-visualization, and (3) animation. Given an image of a scene and a character with a user description, we use these three stages to generate long videos that preserve human identity and provide rich and plausible HSI. Script writing converts a complex text prompt involving a chain of HSI into simple atomic actions that are used in the pre-visualization stage to generate 3D keyframes. To synthesize plausible human interaction poses in 3D keyframes, we utilize pre-trained 2D inpainting diffusion models to generate plausible 2D human interactions based on view canonicalization, which eliminates the need for multi-view fitting in previous works. We then extend these interactions to 3D using robust iterative optimization, informed by contact cues and reasoning from VLMs. Prompted by these 3D keyframes, the pretrained video diffusion models can better generate consistent long videos with plausible dynamics and affordance in a 3D-aware manner. We are the first to synthesize a long video sequence with a chain of HSI actions without training based on the image references of the scene and character. Experiments demonstrate that our method can generate HSI videos that effectively preserve scene content and character identity with plausible human-scene interaction from a single image scene.

Deep neural networks are powerful tools for solving nonlinear problems in science and engineering, but training highly accurate models becomes challenging as problem complexity increases. Non-convex optimization and sensitivity to hyperparameters make consistent performance improvement difficult, and traditional approaches prioritize minimizing mean squared error while overlooking the $L^{\infty}$ norm error that is critical in safety-sensitive applications. To address these challenges, we present HiPreNets, a progressive framework for training high-precision neural networks through sequential residual refinements. Starting from an initial network, each stage trains a refinement network on the normalized residuals of the ensemble so far, systematically reducing both average and worst-case error. A key theme throughout the framework is concentrating training effort on high-error regions of the input domain, which we pursue through complementary techniques including loss function design, adaptive data sampling, localized patching, and boundary-aware training. We validate the framework on benchmark regression problems from the Feynman dataset, where it consistently outperforms standard fully connected networks and reported Kolmogorov-Arnold Networks results, with accuracy approaching machine precision depending on select problems. We further apply the framework to learning the flow map of a 20-dimensional power system ODE, which appears to be the highest dimensional problem studied using this class of multistage methods, achieving substantial reductions in both RMSE and $L^{\infty}$ norm error while enabling a surrogate that predicts system state $238\times$ faster than direct numerical simulation.

Recently, patch deformation-based methods have demonstrated significant effectiveness in multi-view stereo due to their incorporation of deformable and expandable perception for reconstructing textureless areas. However, these methods generally focus on identifying reliable pixel correlations to mitigate matching ambiguity of patch deformation, while neglecting the deformation instability caused by edge-skipping and visibility occlusions, which may cause potential estimation deviations. To address these issues, we propose DVP-MVS++, an innovative approach that synergizes both depth-normal-edge aligned and harmonized cross-view priors for robust and visibility-aware patch deformation. Specifically, to avoid edge-skipping, we first apply DepthPro, Metric3Dv2 and Roberts operator to generate coarse depth maps, normal maps and edge maps, respectively. These maps are then aligned via an erosion-dilation strategy to produce fine-grained homogeneous boundaries for facilitating robust patch deformation. Moreover, we reformulate view selection weights as visibility maps, and then implement both an enhanced cross-view depth reprojection and an area-maximization strategy to help reliably restore visible areas and effectively balance deformed patch, thus acquiring harmonized cross-view priors for visibility-aware patch deformation. Additionally, we obtain geometry consistency by adopting both aggregated normals via view selection and projection depth differences via epipolar lines, and then employ SHIQ for highlight correction to enable geometry consistency with highlight-aware perception, thus improving reconstruction quality during propagation and refinement stage. Evaluation results on ETH3D, Tanks & Temples and Strecha datasets exhibit the state-of-the-art performance and robust generalization capability of our proposed method.

Learned optimizers have been an active research topic over the past decade, with increasing progress toward practical, general-purpose optimizers that can serve as drop-in replacements for widely used methods like Adam. However, recent advances such as VeLO, which was meta-trained for 4000 TPU-months, remain largely inaccessible to the broader community, in part due to their reliance on JAX and the absence of user-friendly packages for independently using the optimizers after meta-training. To address this gap, we introduce PyLO, a PyTorch-based library that brings learned optimizers to the remaining ~70% of machine learning community via the familiar torch.optim.Optimizer interface. Unlike prior work focused on limited-scale academic tasks, our emphasis is on applying learned optimization to real-world large-scale pre-training tasks. Our systems contribution includes CUDA-accelerated implementations of the small fc lopt(Metz et al., 2022a) and VeLO(Metz et al., 2022b) learned optimizers, achieving substantial performance gains, with training throughput on ViT-B/16 (batch size 32) increasing from 39.36 and 49.73 to 205.59 and 191.18 samples per second, respectively. PyLO has the versatility that allows us to easily combine learned optimizers with existing optimization tools such as learning rate schedules and weight decay. When doing so, we discover that learned optimizers can substantially benefit from it. Our code is available at https://github.com/Belilovsky-Lab/pylo

Large language models (LLMs) show strong reasoning abilities but often produce unnecessarily long explanations that reduce efficiency. Although reinforcement learning (RL) has been used to improve reasoning, most methods focus on accuracy and rely on uniform length-based rewards that overlook the differing contributions of individual tokens, often harming correctness. We revisit length optimization in RL through the perspective of token significance. Observing that many chain-of-thought (CoT) tokens contribute little to the final answer, we introduce a significance-aware length reward that selectively penalizes insignificance tokens, reducing redundancy while preserving essential reasoning. We also propose a dynamic length reward that encourages more detailed reasoning early in training and gradually shifts toward conciseness as learning progresses. Integrating these components into standard policy optimization yields a framework that improves both reasoning efficiency and accuracy. Experiments across multiple benchmarks demonstrate substantial reductions in response length while preserving or improving correctness, highlighting the importance of modeling token significance for efficient LLM reasoning.

Reconstructing dynamic hand-object contacts is essential for realistic manipulation in AI character animation, XR, and robotics, yet it remains challenging due to heavy occlusions, complex surface details, and limitations in existing capture techniques. In this paper, we introduce DyTact, a markerless capture method for accurately capturing dynamic contact in hand-object manipulations in a non-intrusive manner. Our approach leverages a dynamic, articulated representation based on 2D Gaussian surfels to model complex manipulations. By binding these surfels to MANO meshes, DyTact harnesses the inductive bias of template models to stabilize and accelerate optimization. A refinement module addresses time-dependent high-frequency deformations, while a contact-guided adaptive sampling strategy selectively increases surfel density in contact regions to handle heavy occlusion. Extensive experiments demonstrate that DyTact not only achieves state-of-the-art dynamic contact estimation accuracy but also significantly improves novel view synthesis quality, all while operating with fast optimization and efficient memory usage. Project Page: https://oliver-cong02.github.io/DyTact.github.io/ .

Modeling propagation is the cornerstone for designing and optimizing next-generation wireless systems, with a particular emphasis on 5G and beyond era. Traditional modeling methods have long relied on statistic-based techniques to characterize propagation behavior across different environments. With the expansion of wireless communication systems, there is a growing demand for methods that guarantee the accuracy and interpretability of modeling. Artificial intelligence (AI)-based techniques, in particular, are increasingly being adopted to overcome this challenge, although the interpretability is not assured with most of these methods. Inspired by recent advancements in AI, this paper proposes a novel approach that accelerates the discovery of path loss models while maintaining interpretability. The proposed method automates the formulation, evaluation, and refinement of the model, facilitating the discovery of the model. We examine two techniques: one based on Deep Symbolic Regression, offering full interpretability, and the second based on Kolmogorov-Arnold Networks, providing two levels of interpretability. Both approaches are evaluated on two synthetic and two real-world datasets. Our results show that Kolmogorov-Arnold Networks achieve the coefficient of determination value R^2 close to 1 with minimal prediction error, while Deep Symbolic Regression generates compact models with moderate accuracy. Moreover, on the selected examples, we demonstrate that automated methods outperform traditional methods, achieving up to 75% reduction in prediction errors, offering accurate and explainable solutions with potential to increase the efficiency of discovering next-generation path loss models.

Although LLM-based agents are proven to master tool orchestration in scientific fields, particularly chemistry, their single-task performance remains limited by underlying tool constraints. To this end, we propose tool amplification, a novel paradigm that enhances the collective capabilities of specialized tools through optimized, dynamic coordination within individual tasks. Instantiating this paradigm, we introduce ChemAmp, a computationally lightweight framework that dynamically treats chemistry tools (e.g., UniMol2, Chemformer) as composable building-block agents. It constructs task-specialized super-agents that transcend atomic tool constraints with limited data ($\leq$10 samples). Our evaluations across four core chemistry tasks molecular design, molecule captioning, reaction prediction, and property prediction demonstrate that ChemAmp outperforms chemistry-specialized models, generalist LLMs, and agent systems with tool orchestration. Critically, this bottom-up construction strategy enables 94\% inference token cost reductions versus vanilla multi-agent systems.

In mathematical reasoning, data selection strategies predominantly rely on static, externally defined metrics, which fail to adapt to the evolving capabilities of models during training. This misalignment limits the efficiency of Supervised Fine-Tuning and Reinforcement Learning. To bridge this gap, we introduce SAI-DPO (Self-Aware Iterative Data Persistent Optimization), a dynamic sampling framework that aligns training data with the model's intrinsic competence. SAI-DPO operationalizes two novel metrics: Knowledge Semantic Alignment for targeting domain weaknesses, and Self-Aware Difficulty, derived from pass rates and reasoning path characteristics, to gauge instance complexity relative to the model's current state. By iteratively recalibrating the data distribution based on real-time feedback, SAI-DPO dynamically aligns training samples with the model's evolving competence, ensuring the data remains strictly relevant to the model's current capability level. Extensive experiments on eight benchmarks (including AIME24 and AMC23) demonstrate that SAI-DPO outperforms static baselines at most nearly 6 points, achieving state-of-the-art efficiency with significantly less data.

Recent advances in reasoning-focused Large Language Models (LLMs) have introduced Chain-of-Thought (CoT) traces - intermediate reasoning steps generated before a final answer. These traces, as in DeepSeek R1, guide inference and train smaller models. A common but under-examined assumption is that these traces are both semantically correct and interpretable to end-users. While intermediate reasoning steps are believed to improve accuracy, we question whether they are actually valid and understandable. To isolate the effect of trace semantics, we design experiments in Question Answering (QA) using rule-based problem decomposition, creating fine-tuning datasets where each problem is paired with either verifiably correct or incorrect traces, while always providing the correct final answer. Trace correctness is evaluated by checking the accuracy of every reasoning sub-step. To assess interpretability, we fine-tune LLMs on three additional trace types: R1 traces, R1 trace summaries, and post-hoc explanations, and conduct a human study with 100 participants rating each type on a Likert scale. We find: (1) Trace correctness does not reliably predict correct final answers - correct traces led to correct solutions in only 28% of test cases, while incorrect traces did not consistently degrade accuracy. (2) Fine-tuning on verbose R1 traces yielded the best model performance, but users rated them least interpretable (3.39 interpretability, 4.59 cognitive load on a 5-point scale), whereas more interpretable decomposed traces did not achieve comparable accuracy. Together, these findings challenge the assumption in question suggesting that researchers and practitioners should decouple model supervision objectives from end-user-facing trace design.

Recently, large reasoning models demonstrate exceptional performance on various tasks. However, reasoning models always consume excessive tokens even for simple queries, leading to resource waste and prolonged user latency. To address this challenge, we propose SelfBudgeter - a self-adaptive reasoning strategy for efficient and controllable reasoning. Specifically, we first train the model to self-estimate the required reasoning budget based on the query. We then introduce budget-guided GRPO for reinforcement learning, which effectively maintains accuracy while reducing output length. Experimental results demonstrate that SelfBudgeter dynamically allocates budgets according to problem complexity, achieving an average response length compression of 61% on math reasoning tasks while maintaining accuracy. Furthermore, SelfBudgeter allows users to see how long generation will take and decide whether to continue or stop. Additionally, users can directly control the reasoning length by setting token budgets upfront.

We introduce softpick, a rectified, not sum-to-one, drop-in replacement for softmax in transformer attention mechanisms that eliminates attention sink and massive activations. Our experiments with 340M and 1.8B parameter models demonstrate that softpick achieves 0\% sink rate consistently. The softpick transformers produce hidden states with significantly lower kurtosis and creates sparse attention maps. Quantized models using softpick outperform softmax on standard benchmarks, with a particularly pronounced advantage at lower bit precisions. Our analysis and discussion shows how softpick has the potential to open new possibilities for quantization, low-precision training, sparsity optimization, pruning, and interpretability. Our code: https://github.com/zaydzuhri/softpick-attention