arXiv:2603.12372v1 Announce Type: new Abstract: Large Reasoning Models (LRMs) have shown remarkable reasoning capabilities, yet they often suffer from overthinking, expending redundant computational steps on simple problems, or underthinking, failing to explore sufficient reasoning paths despite inherent capabilities. These issues lead to inefficiencies and potential inaccuracies, limiting practical deployment in resource-constrained settings. Existing methods to mitigate overthinking, such as suppressing reflective keywords or adjusting reasoning length, may inadvertently induce underthinking, compromising accuracy. Therefore, we propose ReBalance, a training-free framework that achieves efficient reasoning with balanced thinking. ReBalance leverages confidence as a continuous indicator of reasoning dynamics, identifying overthinking through high confidence variance and underthinking via consistent overconfidence. By aggregating hidden states from a small-scale dataset into reasoning mode prototypes, we compute a steering vector to guide LRMs' reasoning trajectories. A dynamic control function modulates this vector's strength and direction based on real-time confidence, pruning redundancy during overthinking, and promoting exploration during underthinking. Extensive experiments conducted on four models ranging from 0.5B to 32B, and across nine benchmarks in math reasoning, general question answering, and coding tasks demonstrate that ReBalance effectively reduces output redundancy while improving accuracy, offering a general, training-free, and plug-and-play strategy for efficient and robust LRM deployment. Code is available at https://github.com/yu-lin-li/ReBalance .

arXiv:2603.12483v1 Announce Type: new Abstract: Across many domains (e.g., IoT, observability, telecommunications, cybersecurity), there is an emerging adoption of conversational data analysis agents that enable users to "talk to your data" to extract insights. Such data analysis agents operate on timeseries data models; e.g., measurements from sensors or events monitoring user clicks and actions in product analytics. We evaluate 6 popular data analysis agents (both open-source and proprietary) on domain-specific data and query types, and find that they fail on stateful and incident-specific queries. We observe two key expressivity gaps in existing evals: domain-customized datasets and domain-specific query types. To enable practitioners in such domains to generate customized and expressive evals for such timeseries data agents, we present AgentFuel. AgentFuel helps domain experts quickly create customized evals to perform end-to-end functional tests. We show that AgentFuel's benchmarks expose key directions for improvement in existing data agent frameworks. We also present anecdotal evidence that using AgentFuel can improve agent performance (e.g., with GEPA). AgentFuel benchmarks are available at https://huggingface.co/datasets/RockfishData/TimeSeriesAgentEvals.

arXiv:2603.12269v1 Announce Type: cross Abstract: Early-exit deep neural networks enable adaptive inference by terminating computation when sufficient confidence is achieved, reducing cost for edge AI accelerators in resource-constrained settings. Existing methods, however, rely on suboptimal exit policies, ignore input difficulty, and optimize thresholds independently. This paper introduces DART (Input-Difficulty-Aware Adaptive Threshold), a framework that overcomes these limitations. DART introduces three key innovations: (1) a lightweight difficulty estimation module that quantifies input complexity with minimal computational overhead, (2) a joint exit policy optimization algorithm based on dynamic programming, and (3) an adaptive coefficient management system. Experiments on diverse DNN benchmarks (AlexNet, ResNet-18, VGG-16) demonstrate that DART achieves up to \textbf{3.3$\times$} speedup, \textbf{5.1$\times$} lower energy, and up to \textbf{42\%} lower average power compared to static networks, while preserving competitive accuracy. Extending DART to Vision Transformers (LeViT) yields power (5.0$\times$) and execution-time (3.6$\times$) gains but also accuracy loss (up to 17 percent), underscoring the need for transformer-specific early-exit mechanisms. We further introduce the Difficulty-Aware Efficiency Score (DAES), a novel multi-objective metric, under which DART achieves up to a 14.8 improvement over baselines, highlighting superior accuracy, efficiency, and robustness trade-offs.

arXiv:2603.12271v1 Announce Type: cross Abstract: LLMs are widely used in knowledge-intensive tasks where the same fact may be revised multiple times within context. Unlike prior work focusing on one-shot updates or single conflicts, multi-update scenarios contain multiple historically valid versions that compete at retrieval, yet remain underexplored. This challenge resembles the AB-AC interference paradigm in cognitive psychology: when the same cue A is successively associated with B and C, the old and new associations compete during retrieval, leading to bias. Inspired by this, we introduce a Dynamic Knowledge Instance (DKI) evaluation framework, modeling multi-updates of the same fact as a cue paired with a sequence of updated values, and assess models via endpoint probing of the earliest (initial) and latest (current) states. Across diverse LLMs, we observe that retrieval bias intensifies as updates increase, earliest-state accuracy stays high while latest-state accuracy drops substantially. Diagnostic analyses of attention, hidden-state similarity, and output logits further reveal that these signals become flatter and weakly discriminative on errors, providing little stable basis for identifying the latest update. Finally, cognitively inspired heuristic intervention strategies yield only modest gains and do not eliminate the bias. Our results reveal a persistent challenge in tracking and following knowledge updates in long contexts.

arXiv:2603.12273v1 Announce Type: cross Abstract: Multi-turn user interactions are among the most abundant data produced by language models, yet we lack effective methods to learn from them. While typically discarded, these interactions often contain useful information: follow-up user messages may indicate that a response was incorrect, failed to follow an instruction, or did not align with the user's preferences. Importantly, language models are already able to make use of this information in context. After observing a user's follow-up, the same model is often able to revise its behavior. We leverage this ability to propose a principled and scalable method for learning directly from user interactions through self-distillation. By conditioning the model on the user's follow-up message and comparing the resulting token distribution with the original policy, we obtain a target for updating the policy that captures how the model's behavior changes in hindsight. We then distill this hindsight distribution back into the current policy. Remarkably, we show that training on real-world user conversations from WildChat improves language models across standard alignment and instruction-following benchmarks, without regressing other capabilities. The same mechanism enables personalization, allowing models to continually adapt to individual users through interaction without explicit feedback. Our results demonstrate that raw user interactions that arise naturally during deployment enable alignment, personalization, and continual adaptation.

arXiv:2603.12278v1 Announce Type: cross Abstract: Diabetic foot ulcers (DFUs) are a severe complication of diabetes, often resulting in significant morbidity. This paper presents a predictive analytics framework utilizing time-series data captured by wearable foot sensors -- specifically NTC thin-film thermocouples for temperature measurement and FlexiForce pressure sensors for plantar load monitoring. Data was collected from healthy subjects walking on an instrumented pathway. Unsupervised machine learning algorithms, Isolation Forest and K-Nearest Neighbors (KNN), were applied to detect anomalies that may indicate early ulcer risk. Through rigorous data preprocessing and targeted feature engineering, physiologic patterns were extracted to identify subtle changes in foot temperature and pressure. Results demonstrate Isolation Forest is sensitive to micro-anomalies, while KNN is effective in flagging extreme deviations, albeit at a higher false-positive rate. Strong correlations between temperature and pressure readings support combined sensor monitoring for improved predictive accuracy. These findings provide a basis for real-time diabetic foot health surveillance, aiming to facilitate earlier intervention and reduce DFU incidence.

arXiv:2603.12288v1 Announce Type: cross Abstract: Tabular machine learning presents a paradox: modern models achieve state-of-the-art performance using high-dimensional (high-D), collinear, error-prone data, defying the "Garbage In, Garbage Out" mantra. To help resolve this, we synthesize principles from Information Theory, Latent Factor Models, and Psychometrics, clarifying that predictive robustness arises not solely from data cleanliness, but from the synergy between data architecture and model capacity. Partitioning predictor-space "noise" into "Predictor Error" and "Structural Uncertainty" (informational deficits from stochastic generative mappings), we prove that leveraging high-D sets of error-prone predictors asymptotically overcomes both types of noise, whereas cleaning a low-D set is fundamentally bounded by Structural Uncertainty. We demonstrate why "Informative Collinearity" (dependencies from shared latent causes) enhances reliability and convergence efficiency, and explain why increased dimensionality reduces the latent inference burden, enabling feasibility with finite samples. To address practical constraints, we propose "Proactive Data-Centric AI" to identify predictors that enable robustness efficiently. We also derive boundaries for Systematic Error Regimes and show why models that absorb "rogue" dependencies can mitigate assumption violations. Linking latent architecture to Benign Overfitting, we offer a first step towards a unified view of robustness to Outcome Error and predictor-space noise, while also delineating when traditional DCAI's focus on label cleaning remains powerful. By redefining data quality from item-level perfection to portfolio-level architecture, we provide a theoretical rationale for "Local Factories" -- learning from live, uncurated enterprise "data swamps" -- supporting a deployment paradigm shift from "Model Transfer" to "Methodology Transfer'' to overcome static generalizability limitations.

arXiv:2603.12296v1 Announce Type: cross Abstract: Deep learning has achieved transformative performance across diverse domains, largely driven by the large-scale, high-quality training data. In contrast, the development of brain-computer interfaces (BCIs) is fundamentally constrained by the limited, heterogeneous, and privacy-sensitive neural recordings. Generating synthetic yet physiologically plausible brain signals has therefore emerged as a compelling way to mitigate data scarcity and enhance model capacity. This survey provides a comprehensive review of brain signal generation for BCIs, covering methodological taxonomies, benchmark experiments, evaluation metrics, and key applications. We systematically categorize existing generative algorithms into four types: knowledge-based, feature-based, model-based, and translation-based approaches. Furthermore, we benchmark existing brain signal generation approaches across four representative BCI paradigms to provide an objective performance comparison. Finally, we discuss the potentials and challenges of current generation approaches and prospect future research on accurate, data-efficient, and privacy-aware BCI systems. The benchmark codebase is publicized at https://github.com/wzwvv/DG4BCI.

arXiv:2603.12298v1 Announce Type: cross Abstract: Activation engineering enables precise control over Large Language Models (LLMs) without the computational cost of fine-tuning. However, existing methods deriving vectors from static activation differences are susceptible to high-dimensional noise and layer-wise semantic drift, often capturing spurious correlations rather than the target intent. To address this, we propose Global Evolutionary Refined Steering (GER-steer), a training-free framework that grounded in the geometric stability of the network's representation evolution. GER-steer exploits this global signal to rectify raw steering vectors, effectively decoupling robust semantic intent from orthogonal artifacts. Extensive evaluations confirm that GER-steer consistently outperforms baselines, delivering superior efficacy and generalization without layer-specific tuning, establishing a universal solution for reliable model alignment.

arXiv:2603.12304v1 Announce Type: cross Abstract: This paper introduces a novel optimization framework that fundamentally integrates the Minimum Description Length (MDL) principle into the training dynamics of deep neural networks. Moving beyond its conventional role as a model selection criterion, we reformulate MDL as an active, adaptive driving force within the optimization process itself. The core of our method is a geometrically-grounded cognitive manifold whose evolution is governed by a \textit{coupled Ricci flow}, enriched with a novel \textit{MDL Drive} term derived from first principles. This drive, modulated by the task-loss gradient, creates a seamless harmony between data fidelity and model simplification, actively compressing the internal representation during training. We establish a comprehensive theoretical foundation, proving key properties including the monotonic decrease of description length (Theorem~\ref{thm:convergence}), a finite number of topological phase transitions via a geometric surgery protocol (Theorems~\ref{thm:surgery}, \ref{thm:ultimate_fate}), and the emergence of universal critical behavior (Theorem~\ref{thm:universality}). Furthermore, we provide a practical, computationally efficient algorithm with $O(N \log N)$ per-iteration complexity (Theorem~\ref{thm:complexity}), alongside guarantees for numerical stability (Theorem~\ref{thm:stability}) and exponential convergence under convexity assumptions (Theorem~\ref{thm:convergence_rate}). Empirical validation on synthetic regression and classification tasks confirms the theoretical predictions, demonstrating the algorithm's efficacy in achieving robust generalization and autonomous model simplification. This work provides a principled path toward more autonomous, generalizable, and interpretable AI systems by unifying geometric deep learning with information-theoretic principles.

arXiv:2603.12305v1 Announce Type: cross Abstract: The ability to understand and reason about cause and effect -- encompassing interventions, counterfactuals, and underlying mechanisms -- is a cornerstone of robust artificial intelligence. While deep learning excels at pattern recognition, it fundamentally lacks a model of causality, making systems brittle under distribution shifts and unable to answer ``what-if'' questions. This paper introduces the \emph{Hierarchical Causal Primitive Dynamic Composition Network (HCP-DCNet)}, a unified framework that bridges continuous physical dynamics with discrete symbolic causal inference. Departing from monolithic representations, HCP-DCNet decomposes causal scenes into reusable, typed \emph{causal primitives} organized into four abstraction layers: physical, functional, event, and rule. A dual-channel routing network dynamically composes these primitives into task-specific, fully differentiable \emph{Causal Execution Graphs (CEGs)}. Crucially, the system employs a \emph{causal-intervention-driven meta-evolution} strategy, enabling autonomous self-improvement through a constrained Markov decision process. We establish rigorous theoretical guarantees, including type-safe composition, routing convergence, and universal approximation of causal dynamics. Extensive experiments across simulated physical and social environments demonstrate that HCP-DCNet significantly outperforms state-of-the-art baselines in causal discovery, counterfactual reasoning, and compositional generalization. This work provides a principled, scalable, and interpretable architecture for building AI systems with human-like causal abstraction and continual self-refinement capabilities.

arXiv:2603.12310v1 Announce Type: cross Abstract: Despite rapid advancements in video generation models, aligning their outputs with complex user intent remains challenging. Existing test-time optimization methods are typically either computationally expensive or require white-box access to model internals. To address this, we present VQQA (Video Quality Question Answering), a unified, multi-agent framework generalizable across diverse input modalities and video generation tasks. By dynamically generating visual questions and using the resulting Vision-Language Model (VLM) critiques as semantic gradients, VQQA replaces traditional, passive evaluation metrics with human-interpretable, actionable feedback. This enables a highly efficient, closed-loop prompt optimization process via a black-box natural language interface. Extensive experiments demonstrate that VQQA effectively isolates and resolves visual artifacts, substantially improving generation quality in just a few refinement steps. Applicable to both text-to-video (T2V) and image-to-video (I2V) tasks, our method achieves absolute improvements of +11.57% on T2V-CompBench and +8.43% on VBench2 over vanilla generation, significantly outperforming state-of-the-art stochastic search and prompt optimization techniques.

arXiv:2603.12324v1 Announce Type: cross Abstract: Connections between statistical mechanics and machine learning have repeatedly proven fruitful, providing insight into optimization, generalization, and representation learning. In this work, we follow this tradition by leveraging results from non-equilibrium thermodynamics to formalize curriculum learning in reinforcement learning (RL). In particular, we propose a geometric framework for RL by interpreting reward parameters as coordinates on a task manifold. We show that, by minimizing the excess thermodynamic work, optimal curricula correspond to geodesics in this task space. As an application of this framework, we provide an algorithm, "MEW" (Minimum Excess Work), to derive a principled schedule for temperature annealing in maximum-entropy RL.

arXiv:2603.12325v1 Announce Type: cross Abstract: Efficient exploration remains a central challenge in reinforcement learning, serving as a useful pretraining objective for data collection, particularly when an external reward function is unavailable. A principled formulation of the exploration problem is to find policies that maximize the entropy of their induced steady-state visitation distribution, thereby encouraging uniform long-run coverage of the state space. Many existing exploration approaches require estimating state visitation frequencies through repeated on-policy rollouts, which can be computationally expensive. In this work, we instead consider an intrinsic average-reward formulation in which the reward is derived from the visitation distribution itself, so that the optimal policy maximizes steady-state entropy. An entropy-regularized version of this objective admits a spectral characterization: the relevant stationary distributions can be computed from the dominant eigenvectors of a problem-dependent transition matrix. This insight leads to a novel algorithm for solving the maximum entropy exploration problem, EVE (EigenVector-based Exploration), which avoids explicit rollouts and distribution estimation, instead computing the solution through iterative updates, similar to a value-based approach. To address the original unregularized objective, we employ a posterior-policy iteration (PPI) approach, which monotonically improves the entropy and converges in value. We prove convergence of EVE under standard assumptions and demonstrate empirically that it efficiently produces policies with high steady-state entropy, achieving competitive exploration performance relative to rollout-based baselines in deterministic grid-world environments.

arXiv:2603.12349v1 Announce Type: cross Abstract: Scientific discovery increasingly relies on AI systems to select candidates for expensive experimental validation, yet no principled, budget-aware evaluation framework exists for comparing selection strategies -- a gap intensified by large language models (LLMs), which generate plausible scientific proposals without reliable downstream evaluation. We introduce the Budget-Sensitive Discovery Score (BSDS), a formally verified metric -- 20 theorems machine-checked by the Lean 4 proof assistant -- that jointly penalizes false discoveries (lambda-weighted FDR) and excessive abstention (gamma-weighted coverage gap) at each budget level. Its budget-averaged form, the Discovery Quality Score (DQS), provides a single summary statistic that no proposer can inflate by performing well at a cherry-picked budget. As a case study, we apply BSDS/DQS to: do LLMs add marginal value to an existing ML pipeline for drug discovery candidate selection? We evaluate 39 proposers -- 11 mechanistic variants, 14 zero-shot LLM configurations, and 14 few-shot LLM configurations -- using SMILES representations on MoleculeNet HIV (41,127 compounds, 3.5% active, 1,000 bootstrap replicates) under both random and scaffold splits. Three findings emerge. First, the simple RF-based Greedy-ML proposer achieves the best DQS (-0.046), outperforming all MLP variants and LLM configurations. Second, no LLM surpasses the Greedy-ML baseline under zero-shot or few-shot evaluation on HIV or Tox21, establishing that LLMs provide no marginal value over an existing trained classifier. Third, the proposer hierarchy generalizes across five MoleculeNet benchmarks spanning 0.18%-46.2% prevalence, a non-drug AV safety domain, and a 9x7 grid of penalty parameters (tau >= 0.636, mean tau = 0.863). The framework applies to any setting where candidates are selected under budget constraints and asymmetric error costs.

arXiv:2603.12433v1 Announce Type: cross Abstract: Model stitching, connecting early layers of one model (source) to later layers of another (target) via a light stitch layer, has served as a probe of representational compatibility. Prior work finds that models trained on the same dataset remain stitchable (negligible accuracy drop) despite different initializations or objectives. We revisit stitching for Vision Foundation Models (VFMs) that vary in objectives, data, and modality mix (e.g., CLIP, DINOv2, SigLIP 2) and ask: Are heterogeneous VFMs stitchable? We introduce a systematic protocol spanning the stitch points, stitch layer families, training losses, and downstream tasks. Three findings emerge. (1) Stitch layer training matters: conventional approaches that match the intermediate features at the stitch point or optimize the task loss end-to-end struggle to retain accuracy, especially at shallow stitch points. (2) With a simple feature-matching loss at the target model's penultimate layer, heterogeneous VFMs become reliably stitchable across vision tasks. (3) For deep stitch points, the stitched model can surpass either constituent model at only a small inference overhead (for the stitch layer). Building on these findings, we further propose the VFM Stitch Tree (VST), which shares early layers across VFMs while retaining their later layers, yielding a controllable accuracy-latency trade-off for multimodal LLMs that often leverage multiple VFMs. Taken together, our study elevates stitching from a diagnostic probe to a practical recipe for integrating complementary VFM strengths and pinpointing where their representations align or diverge.

arXiv:2603.12445v1 Announce Type: cross Abstract: Convolutional Neural Networks have shown promising effectiveness in identifying different types of cancer from radiographs. However, the opaque nature of CNNs makes it difficult to fully understand the way they operate, limiting their assessment to empirical evaluation. Here we study the soundness of the standard practices by which CNNs are evaluated for the purpose of cancer pathology. Thirteen highly used cancer benchmark datasets were analyzed, using four common CNN architectures and different types of cancer, such as melanoma, carcinoma, colorectal cancer, and lung cancer. We compared the accuracy of each model with that of datasets made of cropped segments from the background of the original images that do not contain clinically relevant content. Because the rendered datasets contain no clinical information, the null hypothesis is that the CNNs should provide mere chance-based accuracy when classifying these datasets. The results show that the CNN models provided high accuracy when using the cropped segments, sometimes as high as 93\%, even though they lacked biomedical information. These results show that some CNN architectures are more sensitive to bias than others. The analysis shows that the common practices of machine learning evaluation might lead to unreliable results when applied to cancer pathology. These biases are very difficult to identify, and might mislead researchers as they use available benchmark datasets to test the efficacy of CNN methods.

arXiv:2603.12506v1 Announce Type: cross Abstract: Text-to-Image (T2I) generation is primarily driven by Diffusion Models (DM) which rely on random Gaussian noise. Thus, like playing the slots at a casino, a DM will produce different results given the same user-defined inputs. This imposes a gambler's burden: To perform multiple generation cycles to obtain a satisfactory result. However, even though DMs use stochastic sampling to seed generation, the distribution of generated content quality highly depends on the prompt and the generative ability of a DM with respect to it. To account for this, we propose Na\"ive PAINE for improving the generative quality of Diffusion Models by leveraging T2I preference benchmarks. We directly predict the numerical quality of an image from the initial noise and given prompt. Na\"ive PAINE then selects a handful of quality noises and forwards them to the DM for generation. Further, Na\"ive PAINE provides feedback on the DM generative quality given the prompt and is lightweight enough to seamlessly fit into existing DM pipelines. Experimental results demonstrate that Na\"ive PAINE outperforms existing approaches on several prompt corpus benchmarks.

arXiv:2603.12520v1 Announce Type: cross Abstract: Large language models are often used as judges to score candidate responses, then validated with a single global metric such as correlation with reference labels. This can be misleading when the real deployment task is best-of-n selection within a prompt. In a 5,000-prompt best-of-4 benchmark from Chatbot Arena, a judge with moderate global correlation (r = 0.47) captures only 21.0% of the improvement that perfect selection would achieve over random choice. The gap arises because global agreement is driven largely by prompt-level baseline effects, while selection depends on within-prompt ranking: within-prompt correlation is only r_within = 0.27, and coarse pointwise scoring creates ties in 67% of pairwise comparisons. In a matched-pair best-of-2 audit, explicit pairwise judging recovers much of this lost signal, raising recovery from 21.1% to 61.2%. For judge-based selection, the relevant audit should report within-prompt signal, tie rates, and recovery/top-1 accuracy, not global agreement alone.

arXiv:2603.12529v1 Announce Type: cross Abstract: Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design TERMINATOR, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning TERMINATOR is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train TERMINATOR. Powered by this approach, TERMINATOR achieves significant reductions in CoT lengths of 14%-55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, whilst outperforming current state-of-the-art methods.