Papers

In multi-fidelity optimization, biased approximations of varying costs of the target function are available. This paper studies the problem of optimizing a locally smooth function with a limited budget, where the learner has to make a tradeoff between the cost and the bias of these approximations. We first prove lower bounds for the simple regret under different assumptions on the fidelities, based on a cost-to-bias function. We then present the Kometo algorithm which achieves, with additional logarithmic factors, the same rates without any knowledge of the function smoothness and fidelity assumptions, and improves previously proven guarantees. We finally empirically show that our algorithm outperforms previous multi-fidelity optimization methods without the knowledge of problem-dependent parameters.

Decision-makers rely on weather forecasts to plant crops, manage wildfires, allocate water and energy, and prepare for weather extremes. Today, such forecasts enjoy unprecedented accuracy out to two weeks thanks to steady advances in physics-based dynamical models and data-driven artificial intelligence (AI) models. However, model skill drops precipitously at subseasonal timescales (2 - 6 weeks ahead), due to compounding errors, systemic model biases, and the chaotic nature of the atmosphere. To counter this degradation, we introduce probabilistic bias correction (PBC), a machine learning framework that substantially reduces systematic error by learning to correct historical probabilistic forecasts. When applied to the leading dynamical and AI models from the European Centre for Medium-Range Weather Forecasts (ECMWF), PBC doubles the modest subseasonal skill of the AI Forecasting System and improves the skill of the operationally-debiased dynamical model for 91% of pressure, 92% of temperature, and 98% of precipitation targets. We designed PBC for operational deployment, and, in ECMWF's 2025 real-time forecasting competition, its global forecasts placed first for all weather variables and lead times, outperforming the dynamical models from six operational forecasting centers, an international dynamical multi-model ensemble, ECMWF's AI Forecasting System, and the forecasting systems of 34 teams worldwide. These probabilistic skill gains translate into more accurate prediction of extreme events and have the potential to improve agricultural planning, energy management, and disaster preparedness in vulnerable communities.

Understanding natural and engineered systems often relies on symbolic formulations, such as differential equations, which provide interpretability and transferability beyond black-box models. We introduce Latent Grammar Flow (LGF), a neuro-symbolic generative framework for discovering ordinary differential equations from data. LGF embeds equations as grammar-based representations into a discrete latent space and forces semantically similar equations to be positioned closer together with a behavioural loss. Then, a discrete flow model guides the sampling process to recursively generate candidate equations that best fit the observed data. Domain knowledge and constraints, such as stability, can be either embedded into the rules or used as conditional predictors.

In computer vision and machine learning for geographic data, out-of-domain generalization is a pervasive challenge, arising from uneven global data coverage and distribution shifts across geographic regions. Though models are frequently trained in one region and deployed in another, there is no principled method for determining when this cross-region adaptation will be successful. A well-defined notion of distance between distributions can effectively quantify how different a new target domain is compared to the domains used for model training, which in turn could support model training and deployment decisions. In this paper, we propose a strategy for computing distances between geospatial domains that leverages geographic information with Optimal Transport methods (GeoSpOT). In our experiments, GeoSpOT distances emerge as effective predictors of cross-domain transfer difficulty. We further demonstrate that embeddings from pretrained location encoders provide information comparable to image/text embeddings, despite relying solely on longitude-latitude pairs as input. This allows users to get an approximation of out-of-domain performance for geospatial models, even when the exact downstream task is unknown, or no task-specific data is available. Building on these findings, we show that GeoSpOT distances can preemptively guide data selection and enable predictive tools to analyze regions where a model is likely to underperform.

Data attribution and valuation are critical for understanding data-model synergy for Large Language Models (LLMs), yet existing gradient-based methods suffer from scalability challenges on LLMs. Inspired by human cognition, where decision making relies on a focused readout of relevant memories rather than replaying all pathways, we introduce RISE (Readout Influence Sketching Estimator). Instead of computing and indexing gradients across the entire LLM, RISE focuses on influence hotspots at the output layer, where influence signals concentrate, and the gradient admits a decomposed outer-product form. This enables a dual-channel representation combining a lexical residual channel (RH) and a semantic projected-error channel (GH). Applying CountSketch projections to these channels achieves strong compression while maintaining accurate attribution. Across the OLMo (1B-32B) and Pythia (14M-6.9B) families, RISE reduces index storage by up to 112$\times$ compared to RapidIn and scales to 32B parameters LLM, where gradient-based baselines such as RapidIn and ZO-Inf become memory-infeasible. We evaluate RISE on two paradigms: (1) retrospective attribution, retrieving influential training examples for specific predictions, and (2) prospective valuation, scoring candidate data utility zero-shot. We validate RISE on three tasks: Howdy backdoor data detection, Finance-Medical domain separation, and Brain Rot high-quality data selection. In a closed-loop Brain Rot study, continued pretraining on RISE-selected data yields consistent downstream improvements. Overall, RISE provides a practical and scalable primitive for influence analysis and training-data selection in modern large language models.

Data plays a fundamental role in consolidating markets, services, and products in the digital financial ecosystem. However, the use of real data, especially in the financial context, can lead to privacy risks and access restrictions, affecting institutions, research, and modeling processes. Although not all financial datasets present such limitations, this work proposes the use of deep learning techniques for generating synthetic data applied to cryptocurrency price time series. The approach is based on Conditional Generative Adversarial Networks (CGANs), combining an LSTM-type recurrent generator and an MLP discriminator to produce statistically consistent synthetic data. The experiments consider different crypto-assets and demonstrate that the model is capable of reproducing relevant temporal patterns, preserving market trends and dynamics. The generation of synthetic series through GANs is an efficient alternative for simulating financial data, showing potential for applications such as market behavior analysis and anomaly detection, with lower computational cost compared to more complex generative approaches.

Adapter-based methods have become a cost-effective approach to continual learning (CL) for Large Language Models (LLMs), by sequentially learning a low-rank update matrix for each task. To mitigate catastrophic forgetting, state-of-the-art approaches impose constraints on new adapters with respect to the previous ones, by targeting either subspace or coordinate-wise interference. In this paper, we propose JumpLoRA, a novel framework to adaptively induce sparsity in the Low-Rank Adaptation (LoRA) blocks through the use of JumpReLU gating. The method achieves dynamic parameter isolation, which helps prevent task interference. We demonstrate that our method is highly modular and compatible with LoRA-based CL approaches. Specifically, it significantly boosts the performance of IncLoRA and outperforms the leading state-of-the-art CL method, ELLA.

Large language models (LLMs) increasingly rely on chain-of-thought (CoT) reasoning to solve complex tasks, yet a reasoning trace may merely accompany the final answer rather than causally contribute to it or faithfully reflect how it was reached. We introduce differentiable attention manipulation to train salient reasoning with reinforcement learning – method which we call AtManRL. By optimizing an additive attention mask that identifies the CoT tokens needed to recover the model's answer, we derive a saliency reward and combine it with an outcome reward in GRPO with decoupled advantages. Because both saliency and faithfulness depend strongly on trace length, we evaluate against a sweep of length-penalty baselines that traces the trade-off a pure length penalty can reach, rather than against a single baseline. Across three models (Qwen2.5-3B, Granite-3.3-2B, and Phi-4-mini) and four benchmarks (GSM8K, MMLU, ProofWriter, and QuestBench), AtManRL sits above this trade-off curve by 3.9-4.0pp in saliency, measured as answer changes under mistake injection, and by 5.7-7.7pp in faithfulness, measured as hint verbalization under hint injection, while matching the accuracy of baselines of similar length. These results suggest that training explicitly for saliency is a promising route towards chain-of-thought reasoning that is more faithful to the model's actual decision process.

Accurate prediction of training time in distributed deep learning is crucial for resource allocation, cost estimation, and job scheduling. We observe that the floating-point precision setting is a key determinant of training time, leading to training time variations of ~2.4x over its minimum. However, existing studies on distributed training time prediction rely on static model computation graphs that do not capture precision variations, including mixed precision. According to our experiments, training time prediction without considering precision results in significant prediction errors - reaching up to 147.85% in mean absolute percentage error (MAPE). To address this issue, we propose a precision-aware distributed training time predictor that achieves robust accuracy across diverse precision settings, including mixed precision, with 9.8% MAPE.

We present a dataset and a model for sentiment analysis of German sign language (DGS) fairy tales. First, we perform sentiment analysis for three levels of valence (negative, neutral, positive) on German fairy tales text segments using four large language models (LLMs) and majority voting, reaching an inter-annotator agreement of 0.781 Krippendorff's alpha. Second, we extract face and body motion features from each corresponding DGS video segment using MediaPipe. Finally, we train an explainable model (based on XGBoost) to predict negative, neutral or positive sentiment from video features. Results show an average balanced accuracy of 0.631. A thorough analysis of the most important features reveal that, in addition to eyebrows and mouth motion on the face, also the motion of hips, elbows, and shoulders considerably contribute in the discrimination of the conveyed sentiment, indicating an equal importance of face and body for sentiment communication in sign language.

Accurate molecular property prediction is central to drug discovery, catalysis, and process design, yet real-world applications are often limited by small datasets. Molecular foundation models provide a promising direction by learning transferable molecular representations; however, they typically involve task-specific fine-tuning, require machine learning expertise, and often fail to outperform classical baselines. Tabular foundation models (TFMs) offer a fundamentally different paradigm: they perform predictions through in-context learning, enabling inference without task-specific training. Here, we evaluate TFMs in the low- to medium-data regime across both standardized pharmaceutical benchmarks and chemical engineering datasets. We evaluate both frozen molecular foundation model representations, as well as classical descriptors and fingerprints. Across the benchmarks, the approach shows excellent predictive performance while reducing computational cost, compared to fine-tuning, with these advantages also transferring to practical engineering data settings. In particular, combining TFMs with CheMeleon embeddings yields up to 100% win rates on 30 MoleculeACE tasks, while compact RDKit2d and Mordred descriptors provide strong descriptor-based alternatives. Molecular representation emerges as a key determinant in TFM performance, with molecular foundation model embeddings and 2D descriptor sets both providing substantial gains over classic molecular fingerprints on many tasks. These results suggest that in-context learning with TFMs provides a highly accurate and cost-efficient alternative for property prediction in practical applications.

Multivariate time series classification (MTSC) plays a crucial role in various domains, including biomedical signal analysis and motion monitoring. However, existing approaches, particularly deep learning models, often require high computational resources, making them unsuitable for real-time applications or deployment on low-cost hardware, such as IoT devices and wearable systems. In this paper, we propose the Univariate Channel Fusion (UCF) method to deal with MTSC efficiently. UCF transforms multivariate time series into a univariate representation through simple channel fusion strategies such as the mean, median, or dynamic time warping barycenter. This transformation enables the use of any classifier originally designed for univariate time series, providing a flexible and computationally lightweight alternative to complex models. We evaluate UCF in five case studies covering diverse application domains, including chemical monitoring, brain-computer interfaces, and human activity analysis. The results demonstrate that UCF often outperforms baseline methods and state-of-the-art algorithms tailored for MTSC, while achieving substantial gains in computational efficiency, being particularly effective in problems with high inter-channel correlation.

Practice and extensive exercises are essential in programming education. Intelligent tutoring systems (ITSs) are a viable option to provide individualized hints and advice to programming students even when human tutors are not available. However, prior ITS for programming rarely support the Python programming language, mostly focus on introductory programming, and rarely take recent developments in generative models into account. We aim to establish a novel ITS for Python programming that is highly adaptable, serves both as a teaching and research platform, provides interfaces to plug in hint mechanisms (e.g.\ via large language models), and works inside the particularly challenging regulatory environment of Germany, that is, conforming to the European data protection regulation, the European AI act, and ethical framework of the German Research Foundation. In this paper, we present the description of the current state of the ITS along with future development directions, as well as discuss the challenges and opportunities for improving the system.

We study the sample complexity of learning an $\epsilon$-optimal policy in the Stochastic Shortest Path (SSP) problem. We first derive sample complexity bounds when the learner has access to a generative model. We show that there exists a worst-case SSP instance with $S$ states, $A$ actions, minimum cost $c_{\min}$, and maximum expected cost of the optimal policy over all states $B_{\star}$, where any algorithm requires at least $\Omega(SAB_{\star}^3/(c_{\min}\epsilon^2))$ samples to return an $\epsilon$-optimal policy with high probability. Surprisingly, this implies that whenever $c_{\min} = 0$ an SSP problem may not be learnable, thus revealing that learning in SSPs is strictly harder than in the finite-horizon and discounted settings. We complement this lower bound with an algorithm that matches it, up to logarithmic factors, in the general case, and an algorithm that matches it up to logarithmic factors even when $c_{\min} = 0$, but only under the condition that the optimal policy has a bounded hitting time to the goal state.

We study the problem of learning in zero-sum matrix games with repeated play and bandit feedback. Specifically, we focus on developing uncoupled algorithms that guarantee, without communication between players, the convergence of the last-iterate to a Nash equilibrium. Although the non-bandit case has been studied extensively, this setting has only been explored recently, with a bound of $\mathcal{O}(T^{-1/8})$ on the exploitability gap. We show that, for uncoupled algorithms, guaranteeing convergence of the policy profiles to a Nash equilibrium is detrimental to the performance, with the best attainable rate being $\Omega(T^{-1/4})$ in contrast to the usual $\Omega(T^{-1/2})$ rate for convergence of the average iterates. We then propose two algorithms that achieve this optimal rate up to constant and logarithmic factors. The first algorithm leverages a straightforward trade-off between exploration and exploitation, while the second employs a regularization technique based on a two-step mirror descent approach.

One of the dominant paradigms in self-supervised learning (SSL), illustrated by MoCo or DINO, aims to produce robust representations by capturing features that are insensitive to certain image transformations such as illumination, or geometric changes. This strategy is appropriate when the objective is to recognize objects independently of their appearance. However, it becomes counterproductive as soon as appearance itself constitutes the discriminative signal. In weather analysis, for example, rain streaks, snow granularity, atmospheric scattering, as well as reflections and halos, are not noise: they carry the essential information. In critical applications such as autonomous driving, ignoring these cues is risky, since grip and visibility depend directly on ground conditions and atmospheric conditions. We introduce ST-STORM, a hybrid SSL framework that treats appearance (style) as a semantic modality to be disentangled from content. Our architecture explicitly separates two latent streams, regulated by gating mechanisms. The Content branch aims at a stable semantic representation through a JEPA scheme coupled with a contrastive objective, promoting invariance to appearance variations. In parallel, the Style branch is constrained to capture appearance signatures (textures, contrasts, scattering) through feature prediction and reconstruction under an adversarial constraint. We evaluate ST-STORM on several tasks, including object classification (ImageNet-1K), fine-grained weather characterization, and melanoma detection (ISIC 2024 Challenge). The results show that the Style branch effectively isolates complex appearance phenomena (F1=97% on Multi-Weather and F1=94% on ISIC 2024 with 10% labeled data), without degrading the semantic performance (F1=80% on ImageNet-1K) of the Content branch, and improves the preservation of critical appearance

Traffic forecasting is a challenging spatio-temporal modeling task and a critical component of urban transportation management. Current studies mainly focus on deterministic predictions, with limited considerations on the uncertainty and stochasticity in traffic dynamics. Therefore, this paper proposes an elegant yet universal approach that transforms existing models into probabilistic predictors by replacing only the final output layer with a novel Gaussian Mixture Model (GMM) layer. The modified model requires no changes to the training pipeline and can be trained using only the Negative Log-Likelihood (NLL) loss, without any auxiliary or regularization terms. Experiments on multiple traffic datasets show that our approach generalizes from classic to modern model architectures while preserving deterministic performance. Furthermore, we propose a systematic evaluation procedure based on cumulative distributions and confidence intervals, and demonstrate that our approach is considerably more accurate and informative than unimodal or deterministic baselines. Finally, a more detailed study on a real-world dense urban traffic network is presented to examine the impact of data quality on uncertainty quantification and to show the robustness of our approach under imperfect data conditions. Code available at https://github.com/Weijiang-Xiong/OpenSkyTraffic

Acute Myeloid Leukemia (AML) is one of the most life-threatening type of blood cancers, and its accurate classification is considered and remains a challenging task due to the visual similarity between various cell types. This study addresses the classification of the multiclasses of AML cells Utilizing YOLOv12 deep learning model. We applied two segmentation approaches based on cell and nucleus features, using Hue channel and Otsu thresholding techniques to preprocess the images prior to classification. Our experiments demonstrate that YOLOv12 with Otsu thresholding on cell-based segmentation achieved the highest level of validation and test accuracy, both reaching 99.3%.

Concept Bottleneck Models (CBMs) aim to improve interpretability in Deep Learning by structuring predictions through human-understandable concepts, but they provide no way to verify whether learned concepts align with the human's intended meaning, hurting interpretability. We introduce Prototype-Grounded Concept Models (PGCMs), which ground concepts in learned visual prototypes: image parts that serve as explicit evidence for the concepts. This grounding enables direct inspection of concept semantics and supports targeted human intervention at the prototype level to correct misalignments. Empirically, PGCMs achieve similar predictive performance as state-of-the-art CBMs while substantially improving transparency, interpretability, and intervenability.

The quest for an algorithm that solves an $n\times n$ linear system in $O(n^2)$ time complexity, or $O(n^2 \text{poly}(1/\epsilon))$ when solving up to $\epsilon$ relative error, is a long-standing open problem in numerical linear algebra and theoretical computer science. There are two predominant paradigms for measuring relative error: forward error (i.e., distance from the output to the optimum solution) and backward error (i.e., distance to the nearest problem solved by the output). In most prior studies, convergence of iterative linear system solvers is measured via various notions of forward error, and as a result, depends heavily on the conditioning of the input. Yet, the numerical analysis literature has long advocated for backward error as the more practically relevant notion of approximation. In this work, we show that – surprisingly – the classical and simple Richardson iteration incurs at most $1/k$ (relative) backward error after $k$ iterations on any positive semidefinite (PSD) linear system, irrespective of its condition number. This universal convergence rate implies an $O(n^2/\epsilon)$ complexity algorithm for solving a PSD linear system to $\epsilon$ backward error, and we establish similar or better complexity when using a variety of Krylov solvers beyond Richardson. Then, by directly minimizing backward error over a Krylov subspace, we attain an even faster $O(1/k^2)$ universal rate, and we turn this into an efficient algorithm, MINBERR, with complexity $O(n^2/\sqrt\epsilon)$. Finally, we extend this approach via normal equations to solving general linear systems in $O(n^2\log(n)/\epsilon)$ time complexity. We report strong numerical performance of our algorithms on benchmark problems.

Fine-tuning pre-trained Vision-Language Models (VLMs) for robotic manipulation introduces a fundamental stability-plasticity dilemma: continuous flow-matching action experts backpropagate concentrated, low-rank regression gradients into transformer backbones trained on high-dimensional cross-entropy objectives. This cross-modal gradient asymmetry rapidly degrades pre-trained visual reasoning. Existing solutions either disconnect continuous gradient flow via stop-gradients or constrain updates via LoRA, which restricts update rank but remains directionally blind to semantic corruption; both typically rely on mixed-batch VQA co-training, doubling training compute. We introduce AEGIS (Anchor-Enforced Gradient Isolation System), a buffer-free, layer-wise orthogonal gradient projection framework enabling continuous flow-matching fine-tuning while isolating pre-trained representations from destructive parameter updates. Prior to training, AEGIS estimates per-layer Gaussian activation statistics from pre-training data as a static reference anchor. During fine-tuning, a closed-form Wasserstein-2 transport penalty generates an anchor-restoration gradient through the active computation graph. A sequential dual-backward pass applies layer-wise Gram-Schmidt orthogonalization, projecting task gradients onto the orthogonal complement of the restoration vector during directional conflict. We establish an exact energy preservation bound for layer-wise orthogonal projection, showing that AEGIS sheds only 0.62% of gradient energy empirically while halting cumulative feature drift. On PaliGemma2-3B fine-tuned on the LIBERO manipulation benchmark, AEGIS fully preserves pre-trained Visual Question Answering performance and baseline holdout loss while matching continuous action convergence, without replay buffers, teacher models, or co-training data.

We study discrete k-clustering problems in general metric spaces that are constrained by a combination of two different fairness conditions within the demographic fairness model. Given a metric space (P,d), where every point in P is equipped with a protected attribute, and a number k, the goal is to partition P into k clusters with a designated center each, such that a center-based objective function is minimized and the attributes are fairly distributed with respect to the following two fairness concepts: 1) group fairness: We aim for clusters with balanced numbers of attributes by specifying lower and upper bounds for the desired attribute proportions. 2) diverse center selection: Clusters have natural representatives, i.e., their centers. We ask for a balanced set of representatives by specifying the desired number of centers to choose from each attribute. Dickerson, Esmaeili, Morgenstern and Zhang (2023) denote the combination of these two constraints as doubly constrained fair clustering. They present algorithms whose guarantees depend on the best known approximation factors for either of these problems. Currently, this implies an 8-approximation with a small additive violation on the group fairness constraint. For k-center, we improve this approximation factor to 4 with a small additive violation. This guarantee also depends on the currently best algorithm for DS-fair k-center given by Jones, Nguyen and Nguyen (2020). For k-median and k-means, we propose the first constant-factor approximation algorithms. Our algorithms transform a solution that satisfies diverse center selection into a doubly constrained fair clustering using an LP-based approach. Furthermore, our results are generalizable to other center-selection constraints, such as matroid k-clustering and knapsack constraints.

We introduce the Tan-HWG framework (Hebbian-Wasserstein-Geometry), a geometric theory of Hebbian plasticity in which memory states are modeled as probability measures evolving through Wasserstein minimizing movements. Hebbian learning rules are formalized as Hebbian energies satisfying a sequential stability condition, ensuring well-posed fiberwise JKO updates, optimal-transport realizations, and an energy descent inequality. This variational structure induces a fundamental separation between internal and observable dynamics. Internal memory states evolve along Wasserstein geodesics in a latent curved space, while observable quantities, such as effective synaptic weights, arise through geometric projection maps into external spaces. Simplicial projections recover classical affine schemes (including exponential moving averages and mirror descent), while revealing synaptic competition and pruning as geometric consequences of mass redistribution. Hilbertian projections provide a geometric account of phase alignment and multi-scale coherence. Classical neural networks appear as flat projections of this curved dynamics, while the framework naturally accommodates richer distributional representations, including structural weights and embedding memories, and their spectral extensions in complex internal spaces. Under mild Lipschitz regularity assumptions, including a quasi-stationary "sleep-mode" regime, we establish the existence of continuous-time limit curves. This yields a variational formulation of memory consolidation as a perturbed Wasserstein gradient flow. The framework thus provides a unified geometric foundation for synaptic plasticity, representation dynamics, and context-dependent computation.

The ambulance service is the main transport for diseased or injured people which suffers the same acceleration forces as regular vehicles. These accelerations, caused by the movement of the vehicle, impact the performance of tasks executed by sanitary personnel, which can affect patient survival or recovery time. In this paper, we have trained, validated, and tested a system to assess driving in ambulance services. The proposed system is composed of a sensor node which measures the vehicle vibrations using an accelerometer. It also includes a GPS sensor, a battery, a display, and a speaker. When two possible routes reach the same destination point, the system compares the two routes based on previously classified data and calculates an index and a score. Thus, the index balances the possible routes in terms of time to reach the destination and the vibrations suffered in the patient cabin to recommend the route that minimises those vibrations. Three datasets are used to train, validate, and test the system. Based on an Artificial Neural network (ANN), the classification model is trained with tagged data classified as low, medium, and high vibrations, and 97% accuracy is achieved. Then, the obtained model is validated using data from three routes of another region. Finally, the system is tested in two new scenarios with two possible routes to reach the destination. The results indicate that the route with less vibration is preferred when there are low time differences (less than 6%) between the two possible routes. Nonetheless, with the current weighting factors, the shortest route is preferred when time differences between routes are higher than 20%, regardless of the higher vibrations in the shortest route.

While Large Language Models (LLMs) have achieved strong performance across many NLP tasks, their opaque internal mechanisms hinder trustworthiness and safe deployment. Existing surveys in explainable AI largely focus on post-hoc explanation methods that interpret trained models through external approximations. In contrast, intrinsic interpretability, which builds transparency directly into model architectures and computations, has recently emerged as a promising alternative. This paper presents a systematic review of the recent advances in intrinsic interpretability for LLMs, categorizing existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction. We further discuss open challenges and outline future research directions in this emerging field. The paper list is available at: https://github.com/PKU-PILLAR-Group/Survey-Intrinsic-Interpretability-of-LLMs.

Parallel reasoning enhances Large Reasoning Models (LRMs) but incurs prohibitive costs due to futile paths caused by early errors. To mitigate this, path pruning at the prefix level is essential, yet existing research remains fragmented without a standardized framework. In this work, we propose the first systematic taxonomy of path pruning, categorizing methods by their signal source (internal vs. external) and learnability (learnable vs. non-learnable). This classification reveals the unexplored potential of learnable internal methods, motivating our proposal of STOP (Super TOken for Pruning). Extensive evaluations across LRMs ranging from 1.5B to 20B parameters demonstrate that STOP achieves superior effectiveness and efficiency compared to existing baselines. Furthermore, we rigorously validate the scalability of STOP under varying compute budgets - for instance, boosting GPT-OSS-20B accuracy on AIME25 from 84% to nearly 90% under fixed compute budgets. Finally, we distill our findings into formalized empirical guidelines to facilitate optimal real-world deployment. Code, data and models are available at https://bijiaxihh.github.io/STOP

Post-trained language models produce less varied outputs than their base counterparts. This output diversity collapse undermines inference-time scaling methods that rely on varied samples, and risks homogenizing model outputs on creative and value-laden tasks. Prior work attributes collapse to specific post-training methods, without separating the role of training data composition from the method, or the generation format from the model weights. We trace output diversity through three parallel post-training lineages of Olmo 3, Think (chain-of-thought distillation), Instruct (broad multi-source data), and RL-Zero, across 15 tasks and four text diversity metrics. We find that the location of collapse co-varies with data composition: the Think lineage loses most semantic diversity at supervised fine-tuning, and the effect of DPO is larger in Instruct than in Think. Suppressing chain-of-thought reasoning at inference in Think models drops accuracy on hard tasks, yet leaves answer-level diversity unchanged, showing that the collapse is embedded in the model weights by training data, not imposed by the generation format. Decomposing diversity loss on six verifiable tasks into a quality-control component (removal of incorrect outputs) and a residual component (genuine narrowing among correct outputs) reveals that the split is task-dependent, and Think models retain more correct-answer diversity than Instruct despite collapsing more in aggregate. Our results indicate that diversity collapse is determined during training by data composition and cannot be addressed at inference time alone.

As Large Language Models (LLMs) transition from text processors to autonomous agents, evaluating their social reasoning in embodied multi-agent settings becomes critical. We introduce SocialGrid, an embodied multi-agent environment inspired by Among Us that evaluates LLM agents on planning, task execution, and social reasoning. Our evaluations reveal that even the strongest open model (GPT-OSS-120B) achieves below 60% accuracy in task completion and planning, with agents getting stuck in repetitive behaviors or failing to navigate basic obstacles. Since poor navigation confounds evaluation of social intelligence, SocialGrid offers an optional Planning Oracle to isolate social reasoning from planning deficits. While planning assistance improves task completion, social reasoning remains a bottleneck: agents fail to detect deception at near-random chance regardless of scale, relying on shallow heuristics rather than accumulating behavioral evidence. SocialGrid provides automatic failure analysis and fine-grained metrics, enabling developers to diagnose and improve their agents. We also establish a competitive leaderboard using Elo ratings from adversarial league play.

Interpretable machine learning techniques are becoming essential tools for extracting physical insights from complex quantum data. We build on recent advances in variational autoencoders to demonstrate that such models can learn physically meaningful and interpretable representations from a broad class of unlabeled quantum datasets. From raw measurement data alone, the learned representation reveals rich information about the underlying structure of quantum phase diagrams. We further augment the learning pipeline with symbolic methods, enabling the discovery of compact analytical descriptors that serve as order parameters for the distinct regimes emerging in the learned representations. We demonstrate the framework on experimental Rydberg-atom snapshots, classical shadows of the cluster Ising model, and hybrid discrete-continuous fermionic data, revealing previously unreported phenomena such as a corner-ordering pattern in the Rydberg arrays. These results establish a general framework, which constitutes an important step toward the autonomous discovery of physical descriptors from diverse quantum datasets. All methods are available through qdisc, an open-source Python library designed to make these tools accessible to the broader community.

This paper introduces a robust discrimination method for distinguishing real ship targets from corner-reflector-array jamming with frequency-agile radar. The key idea is to exploit the multidimensional micro-motion signatures that separate rigid ships from non-rigid decoys. From Range-Velocity maps we derive two new hand-crafted descriptors-mean weighted residual (MWR) and complementary contrast factor (CCF) and fuse them with deep features learned by a lightweight CNN. An XGBoost classifier then gives the final decision. Extensive simulations show that the hybrid feature set consistently outperforms state-of-the-art alternatives, confirming the superiority of the proposed approach.