arXiv:2506.11653v3 Announce Type: replace
Abstract: Dataset bias often leads deep learning models to exploit spurious correlations instead of task-relevant signals. We introduce the Standard Anti-Causal Model (SAM), a unifying causal framework that characterizes bias mechanisms and yields a conditional independence criterion for causal stability. Building on this theory, we propose DISCO$_m$ and sDISCO, efficient and scalable estimators of conditional distance correlation that enable independence regularization in gradient-based models. Across six diverse datasets, our methods consistently outperform or are competitive in existing observed bias mitigation approaches, while requiring fewer hyperparameters and scaling seamlessly to multi-bias scenarios. This work bridges causal theory and practical deep learning, providing both a principled foundation and effective tools for robust prediction. Source Code: https://github.com/yakamoz5/DISCO.
Science Journals
arXiv:2506.01467v3 Announce Type: replace
Abstract: Graph generative models perform well on small structured data but struggle to scale to large, complex structures. Hierarchical approaches improve scalability but often ignore node and edge features, which are critical in real-world applications, particularly for hypergraphs that model higher-order relationships. In this paper, we propose FAHNES (feature-aware (hyper)graph generation via next-scale prediction), a hierarchical framework that jointly generates topology and features for graphs and hypergraphs. FAHNES builds multi-scale representations through node coarsening and localized expansion, guided by a novel hierarchical scale encoding that controls granularity and ensures cross-scale consistency. Experiments on synthetic, 3D mesh, and graph point cloud datasets demonstrate competitive or state-of-the-art performance while uniquely scaling to featured large-scale graphs and hypergraphs. Our code is open source
arXiv:2505.20840v2 Announce Type: replace
Abstract: We revisit DropEdge, a data augmentation technique for GNNs which randomly removes edges to expose diverse graph structures during training. While being a promising approach to effectively reduce overfitting on specific connections in the graph, we observe that its potential performance gain in supervised learning tasks is significantly limited. To understand why, we provide a theoretical analysis showing that the limited performance of DropEdge comes from the fundamental limitation that exists in many GNN architectures. Based on this analysis, we propose Aggregation Buffer, a parameter block specifically designed to improve the robustness of GNNs by addressing the limitation of DropEdge. Our method is compatible with any GNN model, and shows consistent performance improvements on multiple datasets. Moreover, our method effectively addresses well-known problems such as degree bias or structural disparity as a unifying solution. Code and datasets are available at https://github.com/dooho00/agg-buffer.
arXiv:2605.30454v1 Announce Type: new
Abstract: Tool-augmented LLM agents are vulnerable to prompt injection: a third party who controls part of the agent's context can plant instructions that the agent then executes as if they came from the user. Current evaluations report a single attack success rate per model on one channel, the tool output and treat that number as the model's vulnerability. But tool descriptions, which the agent reads at every turn before any tool is called, are themselves an injection surface that the attacker can choose instead. We hold the injection payload byte-identical and deliver it through both surfaces across 13 LLMs from six families and four task suites. The same bytes invert in success rate across models: GPT-4.1 is 96 percent vulnerable on tool outputs but only 4 percent on tool descriptions, while GEMINI-3-FLASH shows the mirror pattern at 20 percent and 98 percent. A variance decomposition over 6,830 attempts attributes 0 percent of the variation in attack outcomes to the surface alone, while the model-surface interaction accounts for 16.7 percent. Vulnerability is a property of the pairing, not the channel. The Adaptive Attack Rate, defined as the per-cell maximum over surfaces, exceeds the strongest fixed-surface baseline by +9.1 percentage points on average. Standard prompt-level defenses inherit the same blindspot, reducing tool-output ASR to 10-18 percent while leaving the description channel above 54 percent. Both attack and defense evaluation must report per-surface vulnerability.
arXiv:2505.17607v3 Announce Type: replace
Abstract: Large Language Models (LLMs) display reasoning capabilities over linguistic and symbolic objects but have limited capabilities to directly interpret the continuous numerical outputs of physics simulators, e.g., distances, curvatures, and trajectories that resist discrete tokenisation. Across spatially grounded engineering reasoning tasks, from mechanism design to motion planning, this defines a fundamental gap, which limits the wider application of LLMs within broader geometrical domains, for exmaple interfacing with physics simulators. We propose symbolic intermediaries, compact analytical expressions discovered via symbolic regression, as a structured interface that translates a simulator's numerical traces into a symbolic form, which language models can interpret, compare, and critique while preserving the original geometric semantics. Around this interface we build an agentic coordination-and-refinement loop: a design agent maps natural-language specifications to executable simulation code, a critique agent reasons over the shared symbolic vocabulary, and a revision step turns this feedback into grounded refinement decisions, enabling inference-time generalization without parameter updates. On the MSynth benchmark for planar mechanism synthesis, all three evaluated LLM agents outperform a budget-matched genetic-algorithm baseline by 19-53% (up to 63% lower median error with feedback), and analysis of the critique entries across three model architectures shows that the interface shifts reasoning from generic structural commentary to grounded geometric verification. The principle of translating continuous simulation outputs into symbolic forms generalises to any domain where simulator behaviour must be interpreted linguistically.
arXiv:2505.13356v2 Announce Type: replace
Abstract: Quantum computing has emerged as a promising computational paradigm to address unresolved challenges in the modeling and control of modern power systems. However, most existing studies focus on offline simulations, and a practical framework for validating quantum algorithms in real-time operational environments remains lacking. This study proposes a quantum hardware-in-the-loop framework that integrates a real-time digital simulator with quantum and quantum-inspired hardware to solve combinatorial power flow and optimal power flow formulations under dynamic operating conditions. The proposed framework is validated using the IEEE 9-bus test system and a modified version with integrated solar and wind farms. The results confirm successful integration and convergence within a predefined tolerance. The study also identifies key limitations and challenges, such as limited access to quantum and digital annealers and current scalability limitations, that must be considered in future developments. Nevertheless, the results highlight the potential of quantum computing to significantly enhance the modeling and control of future power systems with high penetration of renewable energy sources.
arXiv:2605.31186v1 Announce Type: new
Abstract: Data streams are nowadays among the most frequently analyzed data structures, with the concept drift posing a major challenge encountered by processing systems. Despite the proposition of numerous solutions to counteract the accuracy degeneration due to concept drift, the scientific community has not yet established a unified framework for evaluating the concept drift detection task. Existing research often relies on classification quality metrics, but these can be affected by multiple factors and may not reliably reflect drift detection quality. In this work, we present an in-depth overview of the relationship between metrics for quantifying drift detection quality and classification performance in synthetic nonstationary data streams. The proposed research studies eight drift detection quality metrics in relation to the classifier's performance across seven synthetic data stream generation tools, additionally considering drift dynamics as a factor. The studies aim to identify the most informative set of drift detection quality metrics and provide a deep understanding of the method's evaluation.
arXiv:2605.31182v1 Announce Type: new
Abstract: We propose a novel anti-twig edge in the honeycomb lattice (HCL) that supports two symmetric high-energy edge states. It is different from the twig edge supporting the zero-energy flat band. Moreover, multiple flat bands are produced by applying a strain to the HCL with a twig edge or an anti-twig edge, and the suppression or enhancement of the high-energy edge state is observed. Under the edge-parallel stretch strain, the high-energy edge state band merges into the bulk band, suppressing the high-energy edge state, while the zero-energy edge state becomes delocalized. On the other hand, under the action of the edge-parallel compression strain, both the high-energy and zero-energy edge states exhibit strong localization. Pseudo-topological protection of the high-energy edge state is explored too. Finally, by reconstructing the anti-twig edge in the HCL, degenerate flat bands and strain-induced multiple flat bands are produced, and topologically protected vacated anti-twig edge states are demonstrated.
arXiv:2605.30861v1 Announce Type: new
Abstract: Post-training for reasoning models typically combines supervised fine-tuning with reinforcement learning from verifiable rewards, most commonly with GRPO. However, this algorithm suffers from sparse rewards, limited exploration, and mode collapse. Building upon recent works on self-distillation, we propose Feedback Distillation, a training method where the model is trained to match, at the token level, its own distribution conditioned on privileged feedback produced by a language model. Feedback Distillation offers token-level supervision and can inject external knowledge. Evaluating our method for Lean4 theorem-proving, we find that Feedback Distillation maintains greater diversity in generated trajectories than GRPO, yielding higher policy entropy and better pass@k scaling. The two methods are complementary: initializing GRPO from a Feedback Distillation checkpoint outperforms either method alone. All in all, our results suggest a promising avenue to improve post-training for complex reasoning.
arXiv:2605.30857v1 Announce Type: new
Abstract: Instruction fine-tuning is employed to enhance the instruction-following ability of large language models (LLMs). As the amount of instruction fine-tuning data increases, selecting the optimal core set becomes particularly important. However, ensuring the diversity of the core set remains a significant challenge. Existing methods predominantly distinguish different training data based on the text features themselves, decoupled from LLMs' own understanding and representation of the data. To address this issue, we propose a Model-Aware Diverse Core Set Selection method, which distinguishes data features based on the neural activation states during LLM inference. This approach serves as an efficient instantiation of coverage-based selection using model-intrinsic activation features to ensure the diversity in the core set. We extensively evaluate our method on six benchmarks that cover five distinct tasks. In our method, the core set selected by the 3B-parameter LLM performs effectively when utilized to fine-tune larger models with 7B, 8B, and 13B parameters. Experimental results on the Alpaca-GPT4 dataset, which comprises 52K instruction-response pairs, show that the core set, sized at 15\% of the original dataset and selected by Llama-3.2-3B-Instruct, achieves an average improvement of 2.5\% when fine-tuning four larger base models compared with training on the full dataset. The experimental results demonstrate that our method enhances model performance on multiple downstream tasks while reducing data requirements.
arXiv:2508.13447v2 Announce Type: replace
Abstract: This work continues the one commenced in \cite{m24}, where the key idea is that individual stances on a social matter can be modeled as positions of particles in physics. Here, we develop an effective thermodynamic framework for the statistical description of collective social systems based on a mechanical representation of individual stances. In this approach, each individual is modeled as a point-like particle evolving in an abstract stance-space, where the dynamics are characterized by position-dependent inertial response functions. The resulting many-particle system is treated within an equilibrium statistical description analogous to the canonical ensemble of statistical mechanics.
Using this framework, we derive the corresponding partition function and introduce emergent macroscopic observables analogous to pressure, volume and temperature, interpreted here as collective statistical properties of the social system. For a particular class of position-dependent mass functions, the resulting equation of state acquires an ideal-gas-like form, suggesting a degree of macroscopic universality despite microscopic heterogeneity among individuals. The formalism is further extended to open social systems through the introduction of a chemical potential associated with population exchange among social groups. This extension naturally allows the treatment of variable particle number and multispecies social configurations within the same equilibrium framework.
The present work is intended as an exploratory phenomenological contribution to sociophysics and complex systems research, aimed at investigating whether collective social behavior may admit effective macroscopic statistical descriptions analogous to those encountered in statistical physics.
arXiv:2605.30846v1 Announce Type: new
Abstract: Object counting remains fragmented across domain-specific datasets and task formulations, despite rapid progress in generalist vision models. Existing counting models are often tailored to scenarios such as crowds, vehicles, cells, crops, or remote-sensing objects, and thus struggle to generalize across categories, visual domains, object scales, and density distributions. In this paper, we study text-guided object counting across domains, where a model takes an image and a natural-language query as input and returns an instance-grounded set of target points whose cardinality gives the count. This formulation unifies category-conditioned counting with interpretable spatial localization. To support this setting, we construct CLOC, a Cross-domain Large-scale Object Counting dataset that reorganizes diverse public data sources into a unified benchmark. CLOC covers six visual domains: General Scene, Remote Sensing, Histopathology, Cellular Microscopy, Agriculture, and Microbiology, with about 220K images, 619 categories, and 15M object instances. Based on CLOC, we propose Count Anything, a generalist model for text-guided object counting. Unlike density-map-based methods, which dominate counting models, Count Anything adopts discrete instance points and performs dual-granularity instance enumeration. A Region-level Sparse Counter provides object-level anchors for large and sparse targets, while a Pixel-level Dense Counter handles small, crowded, and weakly bounded targets via dense point prediction. A point-centric supervision strategy enables learning from heterogeneous annotations, and Complementary Count Fusion combines both counters in a parameter-free manner. Extensive experiments show that Count Anything achieves strong accuracy and multi-domain generalization, outperforming existing open-world counting methods. Code is available at: https://github.com/Mengqi-Lei/count-anything.
arXiv:2509.22335v3 Announce Type: replace
Abstract: We investigate why deep neural networks suffer from loss of plasticity in continual learning, and thus fail to learn new tasks without reinitializing parameters. We show that this failure is preceded by Hessian spectral collapse at new-task initialization, where meaningful curvature directions vanish and gradient descent becomes ineffective. Analyzing a linearized ReLU network, we derive explicit $\epsilon$-rank conditions for successful training and prove that the loss-weighted Gram matrix is spectrally equivalent to the Generalized Gauss-Newton approximation, thereby relating NTK dynamics to Hessian curvature. Targeting spectral collapse directly, we then discuss the Kronecker factored approximation of the Hessian, which motivates two regularization enhancements: maintaining high effective feature rank and applying L2 penalties. Experiments on continual supervised and reinforcement learning tasks confirm that combining these two regularizers effectively preserves plasticity.
arXiv:2601.22068v2 Announce Type: replace
Abstract: Foundation models have become a dominant paradigm in machine learning, achieving remarkable performance across diverse tasks through large-scale pretraining. However, they often yield overconfident, uncalibrated predictions. The standard approach to quantifying epistemic uncertainty are ensembles of multiple independently trained models. But their computational cost scales linearly with ensemble size, making them impractical for large foundation models. We propose Singular Value Ensemble (SVE), a parameter-efficient implicit ensembling method. SVE builds on a simple, but powerful core assumption: namely, that the singular vectors of the weight matrices correspond to meaningful directions in the representation space. If the singular vectors are indeed meaningful (orthogonal) "knowledge directions", then a model ensemble can be obtained by modulating only how strongly each direction contributes to the output. Rather than learning new parameters for each ensemble member, we freeze the singular vectors and only train per-member singular values that rescale the contribution of each direction in that shared knowledge basis. Ensemble diversity emerges naturally during joint training as stochastic initialization and random batch sampling cause different members to converge to different combinations of the same underlying knowledge. SVE performs comparable to an explicit ensemble, while increasing the parameter count of the base model by <1%, making principled uncertainty estimation accessible in resource-constrained settings. We validate SVE on NLP and vision tasks with various different backbones and show that it improves calibration while maintaining predictive accuracy.
arXiv:2504.11972v3 Announce Type: replace
Abstract: Extractive QA tasks are commonly evaluated using Exact Match (EM) and F1-score, but these metrics often fail to reflect true model performance. Recent studies have proposed using large language models (LLMs) as judges (LLM-as-a-judge), yet they often lack comprehensive evaluation across datasets and overlook key factors such as sensitivity to answer types, prompt variations, and self-preference bias. In this work, we conduct a systematic study of LLM-as-a-judge across four extractive QA datasets and various prompt variations, assessing multiple LLM families in both answering and judging roles. Our results show that LLM-as-a-judge judgments correlate much more strongly with human evaluations than EM (0.22) and F1 (0.40), achieving correlations up to 0.85 with open-source models. Further analysis reveals that LLM-as-a-judge performs particularly well on number-related answers but faces challenges with more complex types, such as job titles. Contrary to findings in other NLP tasks, we observe no self-preference bias, even when the same model serves as both QA model and judge. Finally, we find that prompt phrasing has minimal impact, and zero-shot, context-free judging often yields the best evaluation performance.
arXiv:2503.22996v3 Announce Type: replace
Abstract: Sparse Mixture of Experts (SMoE) models scale the capacity of models while maintaining constant computational overhead. SMoE methods fall into two categories: Token Choice, which routes each token to a fixed number of experts, and Expert Choice, which assigns a fixed number of tokens to each expert. However, the use of fixed budgets for tokens or experts causes both approaches to select irrelevant token-expert pairs or overlook critical assignments, which degrades overall performance. To fill that gap, we rethink SMoE from a unified perspective through the lens of linear programming, which provides a general formulation for SMoE models. Furthermore, we introduce Unified Sparse Mixture of Experts (USMoE), a novel framework comprising a unified mechanism and a unified score to overcome these limitations. We provide both theoretical justification and empirical evidence demonstrating USMoE's effectiveness. Extensive evaluations across diverse data settings (clean and corrupted), multiple domains (including texts and vision tasks), and different learning approaches (training-free and training-based) show that USMoE not only delivers significant performance improvements over existing SMoE methods, but also enables more flexible expert selection budgets, reducing inference costs without compromising model performance. Our implementation is publicly available at https://github.com/giangdip2410/USMoE.
arXiv:2605.31210v1 Announce Type: new
Abstract: Crowd movement simulation is essential for pedestrian safety management and facility layout optimization. Data-driven models enhance trajectory prediction accuracy under Euclidean metrics, yet they suffer from excessively high collision rates, especially in bidirectional and multidirectional flows. In this paper, we establish a novel data-driven crowd simulation model that incorporates the pedestrian collision mechanism into the loss function to reduce collisions. A new lateral-acceleration-based collision loss function and a Voronoi-based motion feature extraction approach are proposed. The model is based on a Generative Adversarial Network (GAN) architecture and is termed CPGAN (Collision-Penalized GAN). We evaluate CPGAN in bidirectional flow scenarios, which involve frequent collision avoidance behaviors. Results show that the proposed lateral-acceleration-based collision loss significantly reduces opposite-direction pedestrian collision rates to levels comparable with controlled experiments. CPGAN effectively simulates bidirectional flow, reproducing lane formation and N-t curves. The research outcomes can provide inspiration for integrating pedestrian dynamics mechanisms into loss functions in data-driven crowd simulation.
arXiv:2605.31121v1 Announce Type: new
Abstract: Outdoor vision-language navigation (VLN) in long-range, open-world environments is frequently disrupted by semantic-cue interruptions, where informative goal cues become sparse, occluded, or leave the field of view. Once such cues disappear, agents enter a cue-free phase and often degrade into backtracking, oscillatory headings, or aimless exploration. While memory-based methods attempt to bridge these gaps, they often fail under traversability-driven detours: the remembered cue direction may be infeasible, forcing detours that prolong cue-free phases and gradually render robot-centric cues stale and implicit histories blurred. This makes traversability a stability condition for maintaining goal-directed guidance, rather than merely a local safety concern.
We propose a unified outdoor VLN framework that survives semantic-cue interruptions by maintaining traversability-consistent executable guidance throughout prolonged cue-free phases. Specifically, our method extracts semantic bearings from visibility-gated goal or exploration cues and grounds them into executable headings using a real-time near-field traversability profile, providing goal-consistent feasible guidance beyond reject-only safety filtering. To prevent guidance degradation during detours, we lift intermittent 2D evidence into a world-aligned 3D cue memory with an uncertainty-aware readout mechanism, ensuring guidance remains continuously reachable and stable as the robot moves.
We evaluate the framework on quadrupedal and wheeled platforms over 600--1000 m routes. Our method improves simulation success rate by over 10 percentage points over the strongest baseline and achieves a real-world success rate of 40%, compared to 17.5% for the strongest baseline, with substantially higher robustness during prolonged cue-free intervals.
arXiv:2605.30568v1 Announce Type: new
Abstract: LLM-as-a-Judge is a scalable alternative to human evaluation, yet existing rubric-based methods rely on human-annotated data such as reference answers or expert-crafted rubrics. We propose to automatically generate fine-grained evaluation rubrics without any human annotation. Our training-free method generates rubrics at dataset-specific and instance-specific granularities, achieving performance competitive with existing methods across four benchmarks. We further present a method that iteratively fine-tunes a rubric generator model via meta-judge reward signals. The fine-tuned generator outperforms all existing baselines in both pairwise and pointwise evaluation. Notably, a fine-tuned 14B rubric generator outperforms a much larger proprietary model at rubric generation, showing the effectiveness of our fine-tuning strategy.
arXiv:2605.31105v1 Announce Type: new
Abstract: Large language models (LLMs) with extended context lengths rely on the key-value (KV) cache to support attention over prior tokens. However, maintaining the KV cache incurs substantial memory overhead, motivating KV-cache compression methods that enforce a fixed budget through eviction and merging. Modern eviction methods increasingly adopt span-based retention because preserving contiguous spans is empirically effective and better preserves semantic coherence. Yet, when combined with post-eviction merging, span-based retention concentrates merges onto a small set of span-boundary carrier tokens, producing a highly imbalanced merge pattern that exacerbates over-merging and increases information loss. To address this imbalance, we propose GRKV (Global Regression for KV Cache), a training-free KV-cache merging method that directly minimizes the discrepancy between compressed-cache and full-cache attention outputs. GRKV uses ridge-regression-based merge steps to distribute information from evicted tokens across retained tokens, while regularizing the updates to prevent over-smoothing. Across the LongBench and RULER long-context benchmarks, GRKV is the only merging method that improves overall performance with minimal overhead.
arXiv:2605.31117v1 Announce Type: new
Abstract: The long-term accessibility and reusability of scholarly knowledge is a central concern of Open Science. Research and infrastructure development in this area have so far focused predominantly on how traditional scientific outputs, such as journal articles, monographs, and conference proceedings, can be preserved and made openly available over time. Alternative forms of scholarly communication such as scholarly blogs, by contrast, have received comparatively little attention, even though they have become an established medium for disseminating research and for fostering dialogue within academia and with the wider public. The lack of preservation of scholarly blogs puts them at a risk of information loss, which poses the threat of leaving a gap in the scholarly record. Prior research has examined how blogs are integrated into information infrastructures and what requirements scholarly bloggers have for an information infrastructure that ensures long-term access to their blogs. What is needed now are recommendations for the implementation of these results into the library practice. Based on a convergent mixed-methods design that merges a quantitative analysis of 866 German scholarly blogs, a qualitative interview study with 13 scholarly bloggers, and an open, participatory review process with the scholarly blogging community, we propose a catalog of requirements for the integration of scholarly blogs into information infrastructures in order to ensure their long-term accessibility, reusability and citeability.
arXiv:2605.31111v1 Announce Type: new
Abstract: Joint-Embedding Predictive Architectures (JEPAs) learn compact latent world models by predicting future embeddings, but no single coordinate of the latent is designated to encode task progression. We carve the JEPA latent into two orthogonal subspaces with disjoint roles: a low-dimensional progression subspace shaped by a cosine-margin triplet loss, and a high-dimensional content subspace regularised by the existing SIGReg objective of LeWM. We prove that the two anti-collapse forces act on disjoint coordinates, so they compose additively rather than competing on the same dimensions. Our method, SD-JEPA improves over the LeWM baseline on the majority of its control benchmarks at matched compute, and outperforms the strongest non-LeWM JEPA baseline on Push-T; a subspace-ablation falsifier confirms the split is the load-bearing ingredient. Beyond planning, the resulting 1-D angular progression coordinate functions as a scene-aware compass on the latent. It advances with task progress, regresses when the agent backtracks, and under controlled perturbations both spikes and relocalises to a semantically appropriate new task-phase sector, separating the moment of surprise from its meaning in a way that prediction-error scalars cannot. Three quantitative tests back this up: $|\Delta\theta_t|$ outperforms the standard latent-prediction-error surprise at localising semantic events on 40 held-out cube episodes by up to +0.18 pooled AUROC (97.5% per-episode win rate at $\pm 1$-step tolerance); a within-episode linear probe across all four environments (40 episodes per env) shows the 8-dimensional progression subspace (4.2% of the latent) explains 72-95% of task-progress variance..
arXiv:2605.30812v1 Announce Type: new
Abstract: Accurately modeling the macroscopic dynamics of high-dimensional microscopic systems is of broad interest across the sciences. Many data-driven approaches learn a low-dimensional latent state through an autoencoder trained for pointwise input reconstruction. These methods typically assume a fixed ordering of microscopic degrees of freedom in the input. However, in many settings, such as particle systems, the microscopic state is inherently unordered. This motivates an autoencoder framework that learns permutation-invariant latent representations. To this end, we adopt a permutation-invariant encoder and design the decoder to reconstruct the mass distribution centered at the observed points rather than per-sample reconstruction. We then jointly learn the macroscopic dynamics of the observables together with the latent states. We demonstrate the effectiveness and robustness of the proposed method across a range of microscopic settings, including learning the energy dynamics in interacting particle systems, predicting mixing dynamics in Lennard-Jones fluids, and modeling the stretching dynamics from video data of polymers moving in an elongational force field.
arXiv:2605.30790v1 Announce Type: new
Abstract: Retrieval-Augmented Generation (RAG) supplements a language model's input with retrieved documents, yet most RAG pipelines inherit retrieval components designed for human readers. How retrieved content should be represented when the consumer is a large language model (LLM) rather than a human is less well understood. Recent work has proposed transformations of retrieved content and identified properties that affect generation, but each examines a single transformation or property in isolation, leaving open which features of a document's representation matter most. We address this with a controlled comparison: holding retrieval fixed, we vary only the representation of retrieved documents, comparing an original baseline against thirteen transformations spanning selection, summarisation, and reformulation, in query-dependent and query-independent variants. Across these fourteen representations we measure question-answering accuracy for four generators, and for each representation we also measure answer retention: whether a known answer-bearing document still supports its answer after transformation. We find that answer retention is the primary determinant of generator accuracy; notably, when retention is high, a representation's wording, structure, length, and query-dependence have limited effect. This suggests that accuracy gains attributed to specific mechanisms in prior work may be partly explained by how well those mechanisms preserve answer-bearing content, an attribution that cannot be settled without controlling for retention.
arXiv:2605.30646v1 Announce Type: new
Abstract: Large Language Models (LLMs) are increasingly used in clinical applications. However, their behavior remains highly sensitive to subtle linguistic variations, such as rephrasing or syntactic variation. This sensitivity poses risks in safety-critical healthcare settings, where semantically equivalent inputs should produce consistent predictions. However, a key challenge is to ensure that prompt variations truly preserve clinical meaning, as embedding-based similarity metrics often fail to capture distinctions involving negation, temporality, or severity. To address this limitation, we propose a semantic verification framework based on Natural Language Inference (NLI) to filter meaning-preserving prompt variations, which are further refined using an LLM-as-a-judge and audited by a clinical expert. In addition, we introduce three metrics to quantify model sensitivity: MeaningPreserving Variation Sensitivity (MVS), confidence variation (\Delta C), and Worst-Case Instability (WCI). We evaluate 16 open-source general-purpose (GP) and medical LLMs within the same model families and parameter scales, using reformulated prompts derived from the DiagnosisQA and MedQA datasets. Our results demonstrate that robustness differences between domain-specific (DS) models are mixed and highly model-dependent, i.e., domain specialization does not consistently improve or reduce robustness to meaning-preserving prompt reformulations. Several DS models rank among the most robust (when compared with GP counterparts), and strong GP baselines remain competitive as well.