Computation and Language
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
Wednesday 19 August
AI hallucinations - that is, outputs which are made up, cannot be verified, or contradict the source material - are generally regarded as an engineering flaw to be dealt with. This paper contends that they also have philosophical significance when it comes to the question of machine consciousness. We examine the known causes of hallucinations in large language models - such as source-target divergence, discrepancies between training and inference, and overfitting - and we present two empirical investigations. In the first, we apply successive generations of the GPT model to ambiguous factual questions under different temperature settings, finding that higher temperatures result in plausible but incorrect answers while lower temperatures lead to factually accurate ones. The sampling parameters that cause a model to seem creative or spontaneous and thus more likely to pass behavioral tests of intelligence are the same ones that increase its hallucination rate. In the second, we look at an encoder-only model that has been trained on encyclopedic data and which answers questions of the same type factually and without embellishment, indicating that hallucinations are due to exposure to subjective and socially diverse training data rather than to the development of any cognitive ability. Using references to Turing, Searle's Chinese Room, the frame problem, and the cybernetic tradition of Wiener and Ashby, we claim that a model's self-reports of emotion or sentience come within the definition of hallucination, and that any future occurrence of machine consciousness might remain epistemically inaccessible since it would be indistinguishable from a sufficiently advanced hallucination.
Formal Languages and Automata
A strengthening of the MCFL-ness of $O_2$
In the last years, a number of proofs of the fact that $O_2$ is a multiple context-free grammar (MCFG) were given. Such results can be exploited in the fields of both computational linguistics and of computational algebra. Here, we focus on a recent such proof spelled in terms of factorizations of string tuples, and give a new result with a stronger characterization of such factorizations than in existing theorems.
Read the full abstract
Machine Learning
Many Optimizers But Only One Training Path: Repeated Resampling for Adaptive Optimizer Selection
An optimizer is usually chosen before training a deep neural network and then kept fixed. Treating optimizer choice as a hyperparameter could boost performance, but it requires several complete training runs and discards all but the winner. Repeated Optimizer Resampling (ROR) instead searches during one evolving run. Every $b$ epochs, each candidate optimizer scouts from the current model weights for $s$ epochs. The best scout continues for the remaining $b-s$ epochs, and that completed segment becomes the new incumbent if it improves the validation objective. This design allows the preferred optimizer to change as training progresses. We compare two variants of ROR on MNIST, Fashion-MNIST, and two motor insurance claim-count models. Nine fixed optimizers and both ROR variants are evaluated with the same ten seeds. One-epoch ROR uses 24% to 35% of the aggregate training needed to identify the best fixed optimizer exhaustively and remains close to that optimizer on all four tasks. These results support short scouting as a practical way to search over optimizers without completing every candidate run.
Read the full abstract
Numerical Analysis
Some uniform error analysis for gKdV equations in dispersionless limit regime before dispersive shock
This work establishes uniform error estimates for classical numerical schemes applied to the generalized Korteweg-de Vries (gKdV) equation in the dispersionless limit regime, specifically before the development of dispersive shock wave (DSW). We analyze the Crank-Nicolson (CN) and a Lawson-type Runge-Kutta (Lawson-RK) methods, when discretized in space via the Fourier pseudo-spectral method. We prove that both fully discrete schemes achieve optimal second-order temporal accuracy and spectral spatial accuracy, with error constants being uniform in the vanishing dispersion parameter $\varepsilon$. The analysis also addresses the unique solvability of the CN scheme in the dispersionless limit regime. Theoretical findings are supported by numerical experiments, demonstrating the $\varepsilon$-uniform accuracy and the necessity of step size restriction for CN. The study in general validates the classical methods for solving the dispersionless gKdV equation before DSW.
Read the full abstract
Machine Learning
Tensor Field Models
This paper introduces Tensor Field Models (TFMs), realization-level Mathematical Structures in which a learned Operator maps a product of admissible component-section families to a prescribed family of time-dependent tangent sections on a Generative State Manifold. Analytic and dynamical restrictions are encoded through the choice of admissible families rather than imposed by the root definition. Constructed, component-separable, and Tensor Bundle TFMs provide structured refinements of this common object. In the conditional realizations considered here, a structured condition $c=(c_1,\ldots,c_n)$ is mapped componentwise to a reusable collection $\mathbf H_c=(H_{c_1}^{(1)},\ldots,H_{c_n}^{(n)})$. In the architectures evaluated here, the component representations remain distinct and are combined only by the Field Operator to produce the generated Vector Field. All learned models are trained using Flow Matching. Experiments show that TFMs can improve performance and that amortized sampling enabled by reusable condition representations can accelerate generation.
Read the full abstract
Machine Learning
Forgetting, plasticity, and co-observation: a third facet of continual learning
Efficient continual learning remains a fundamental challenge for deep neural networks. While catastrophic forgetting and loss of plasticity are widely considered the primary obstacles to overcome, we show that these two issues cannot fully explain the performance gap between naive sequential training and offline joint training. In this paper, we highlight data co-observation as a distinct factor influencing continual learning performance. By decoupling the constraints of separate data access from stability and plasticity, we systematically investigate the representational benefits gained by observing training data together. Empirically, we demonstrate a consistent performance difference between joint and separate training across both supervised and self-supervised paradigms in generic data-incremental "chunking" scenarios, whilst mitigating forgetting and controlling for plasticity. Our findings indicate that simultaneous observation of training data (co-observation) yields benefits to the learner's generalization that extend well beyond mere knowledge retention, and that this effect does not require a specific continual distribution shift. Furthermore, we contextualize prominent continual learning mechanisms through this lens: while distillation-based approaches act only as effective knowledge retention mechanisms, our results suggest that the empirical success of memory replay goes beyond the mitigation of forgetting, actively reintroducing the benefits of data co-observation into the learning process.
Read the full abstract
Human-Computer Interaction
A revised framework for the assessment of psychological safety in autonomous vehicles
Despite recent technological progress in the development of autonomous vehicles (AVs), their societal acceptability remains a subject of debate as recent research findings point to psychological roadblocks. Concerns arise not only for physical safety but also for potential psychological risks resulting from human interaction with AVs. Psychological concepts such as trust, and perceived safety are well-studied in this context and are found to be determinant factors for the intention to use AVs. Unfortunately, there has been no formalization of the mechanism by which human interaction with AVs may lead to psychological hazards, threatening trust, perceived safety, and acceptability. Furthermore, there has been little prior research that conceptualizes the severity of psychological risk in AVs, and there are no clear guidelines for a systems designer on how to assess and address psychological risk in the AV development context. To address these limitations, this paper extends a theoretical framework for AV psychological safety risk assessment based on an early proposal. The proposed framework consists of an extended risk model for psychological safety including all the key concepts related to psychological safety in AVs, and an assessment method based on the Systems-Theoretic Accident Model and Processes (STAMP). We demonstrate the usefulness of the theoretical framework through a highly automated AV use case scenario, uncovering factors which may lead to psychological risk for an occupant. The use cases provide examples of how to use the framework to extensively evaluate psychological risk and determine vehicle behaviour that could lead to this risk. By developing a theoretical framework for AV psychological safety risk assessment, we provide a foundation and a method that were previously lacking to better enable responsible AV development regarding psychological safety.
Read the full abstract
Social and Information Networks
Comparing Probabilistic Influence-Spreading Centralities to Commonly-Used Centrality Measures in Directed and Weighted Networks
The Influence-Spreading Model (ISM) introduces three probabilistic centrality measures: out-centrality, in-centrality, and ISM betweenness centrality. Out-centrality measures the average probability that a node influences others, while in-centrality measures the average probability that others influence a node. ISM betweenness centrality measures the change in total probabilistic influence when a node is removed. These measures depend on edge transmission probabilities and allow walks up to a specified maximum length. We compare the ISM centrality measures to commonly used weighted variants of out-degree, in-degree, closeness, shortest-path betweenness, and Katz centrality in directed, weighted networks using four real-world online social networks and nine synthetic networks generated by Erdős-Rényi, navigable small-world, and directed scale-free models. For the synthetic networks, the edge probabilities are drawn from three beta distributions. We evaluate the similarity in centrality values and their ranking using Pearson correlation and Spearman's rank correlation coefficients. Results show strong correlations between the ISM out-centrality and weighted out-degree and outward Katz centrality, particularly for low edge probabilities. Conversely, relationships between the ISM in-centrality and other measures vary with network topology, sometimes yielding negative correlations. Correlations between the ISM betweenness and the shortest-path betweenness are also topology-dependent and weaken as alternative influence paths become more relevant. Overall, standard centrality measures can approximate the influence of broadcasting influence but often miss the nuances of receiving influence and probabilistic intermediary roles.
Read the full abstract
Information Theory
A Framework for Enterprise Network Dimensioning
We study radio node (RN) placement for indoor enterprise networks. Using stochastic geometry (SG), we derive the meta-distribution (MD) of the SINR for a test user equipment (UE), with and without cooperation from outdoor macro base stations (MBSs), and compare these results with an integer linear programming (ILP) approach. SG provides an estimate of the required number of RNs but not their locations, while ILP can yield inaccurate local optima and requires high computational power. To address this, we investigate clustering-based algorithms for initializing RN locations using UE location distributions. Along with standard methods, we propose a weighted $k$-harmonic means (WKHM) clustering strategy tailored to maximize SINR. We then introduce a constrained sequential minimum cut algorithm, SeqMinCut, to merge multiple RNs into larger cells and further improve SINR. This is the first work that integrates SG-based statistical analysis, optimization, and clustering to obtain system design insights, dimensioning rules, and planning strategies for enterprise 5G.
Read the full abstract
Computation and Language
Decomposing Wrong-Consensus Agreement in LLM Self-Consistency
Agreement among repeated samples of a language model is routinely read as evidence about answer reliability, yet wrong answers can agree just as strongly as right ones. This paper asks what information wrong-consensus agreement actually contains, and answers with a quantitative decomposition. A pluralistic agreement index Gamma, normalized by the reference scale d=(1-p)/(C-1), is split into a mechanical component (agreement delivered by a per-case answer preference alone) and a preference-unexplained residual. The mechanical reference is leak-free: each case's preference and accuracy are estimated from its other runs only. On public GPT-4.1 per-run data, coverage phi (the mechanical/empirical ratio) shows a benchmark-associated direction: 0.81-0.93 on multiple-choice GPQA-Diamond against 0.59-0.78 on open-domain AIME, where a residual of 1.54-2.80 Gamma units survives, more than absorbed by a calibrated run-level preference-heterogeneity reference. A controlled replication under one fixed protocol (four runs per question, K=32 votes) on five open-weights checkpoints (Qwen3.5-9B/122B, Qwen3.8-27B, Gemma4-26B/31B) finds near-complete mechanical coverage in all ten cells (phi approximately 1, with a small overshoot consistent with a quantified finite-donor plug-in bias), robust to a two-run design; the largest cell (qwen3.5-122b, p=0.222) sits inside the GPT-4.1 AIME accuracy range and still saturates (phi=1.041). A cross-system contrast at comparable aggregate accuracy contrasts near-complete mechanical agreement in the open-weights models against a larger preference-unexplained residual in the frontier family. This contrast is confounded with sampling protocol by design. Agreement is graded evidence, not certification. No new voting method is proposed; code and evidence are committed.
Read the full abstract
Robotics
Dream2Reward: Transition-Alignment Reward Models from Positive Demonstrations for Robotic Manipulation
Learning robotic policies requires dense rewards that remain informative when behavior departs from successful demonstrations. Progress-based rewards estimate how far an observation has advanced along a nominal successful trajectory, but may remain high after an incorrect transition. We introduce Dream2Reward, which learns a language-conditioned successful latent transition field from positive demonstrations. Given the visual history up to a transition start, the model predicts the latent displacement associated with successful execution and scores the observed displacement through signed directional and symmetric magnitude agreement. This transition-level comparison penalizes wrong-direction, overshooting, and stagnant motion even when the resulting observation appears to show progress. Dream2Reward requires no failure annotations, progress labels, or synthetic negatives, and produces a dense causal reward. Across mechanism diagnostics and shared-trajectory evaluations, it provides stronger success-failure separation and more informative feedback on low-quality behavior than progress-based alternatives. Across online and offline policy learning, the same frozen reward model reduces reward hacking and supports stronger downstream performance, including in real-robot manipulation. These results show that comparing realized motion with predicted successful change provides an effective way to convert positive demonstrations into dense rewards for robot learning.
Read the full abstract
Human-Computer Interaction
Scoring and Gamification to Encourage Sustainable Use of Compute Clusters
The environmental cost of computing continues to grow, yet behaviour change remains limited. We present a composite sustainability score integrating average carbon intensity, resource utilisation, and embodied emissions into a single 0-100 metric designed for gamified feedback. Each component rewards a different dimension of sustainable behaviour: carbon-aware workload shifting, high resource utilisation, and selecting hardware that is commonly underutilised. This scoring system is built into an existing cluster management interface and underpins three dashboard conditions: raw metrics, composite score, and a gamified tree visualisation, which we are planning to evaluate in a 12-week within-subjects study with approximately 35 researchers. Furthermore, we open the discussion on the challenge of defining computational work `goodness' in the context of sustainability scores.
Read the full abstract
Systems and Control · Cross-listed from math.OC
Mitigating Regional Traffic Congestion via School Start Time Scheduling: A Bilevel Alternating Optimization Approach
This paper addresses morning commute congestion caused by concentrated school-related trips in urban networks. We propose a bi-level optimization framework for regulating school start times in a multi-region urban network characterized by Macroscopic Fundamental Diagrams (MFDs), explicitly coupling system-level regulation with multi-class user-equilibrium-based departure-time choices. The Upper-Level problem jointly minimizes total time spent and deviations from current school schedules, while the Lower-Level problem models commuter behavior through a deterministic dynamic multi-class user equilibrium formulation incorporating alpha-beta-gamma preferences for travel time, earliness, and lateness costs. To address the computational challenges arising from the bilevel structure, non-convex traffic dynamics, and endogenous demand responses, an iterative algorithm alternating between the Upper- and Lower-Level problems is developed. The Upper-Level problem is approximated through a formulation solvable with standard mathematical programming solvers, while an iterative algorithm provides an approximate solution to the Lower-Level equilibrium problem. Numerical results demonstrate substantial congestion reductions and characterize the trade-off between school start-time flexibility and traffic efficiency. Sensitivity analyses further examine the effects of MFD uncertainty and scheduling preferences.
Read the full abstract
Computational Complexity
On the quantum communication complexity of total functions
We present a total function with a polylogarithmic two-message quantum protocol, whereas every randomised protocol, even with arbitrarily many rounds, requires polynomial communication.
Machine Learning
A Real-Time Tsetlin Machine-based Non-intrusive Load Monitoring System on MCUs
Non-Intrusive Load Monitoring (NILM) systems estimate individual appliance energy consumption from a single aggregate meter, without requiring separate sensors for each device. By installing a single meter that measures a building's total electricity consumption, NILM algorithms can determine the active status of each appliance. However, traditional NILM systems use computationally intensive optimization algorithms to process offline data, limiting their capability for on-device deployment, where sensitive household data must be processed locally. This paper proposes a Tsetlin Machine (TM)-based NILM framework, targeting real-time applications on resource-constrained microcontrollers (MCUs), enabling privacy-preserving edge deployment. The problem is reformulated as a classification task, and the proposed approach achieves an average precision of 90% and recall of 96% for two-appliance classification, and 77% precision and 80% recall for four appliances on the REDD dataset. The trained model occupies only 17 KB of flash memory and achieves an inference latency of 0.43 ms on an ESP32, demonstrating its suitability for NILM applications on MCUs.
Read the full abstract
Information Retrieval
SIDScope: A Diagnostic Resource for Semantic-ID Interfaces in Generative Recommendation
Semantic ID tokenizers assign code sequences to catalog items, which generative recommenders predict and resolve back to items. Reusing a released mapping requires checking its code organization, item resolution, and compatibility with an existing model after a refresh. SIDScope provides these checks through source traced adapters, mapping diagnostics, candidate retrieval tests, and generated trace analysis. Its inventory contains nine exports from seven tokenizer families across Amazon and Yelp, including eight executable routes and one auditable snapshot. The analyses establish three distinctions. First, code allocation and item addressability differ: in GRID, tail items occupy 95.7 distinct codes per 100 items, yet only 82.2 percent have globally unique codes. Second, prefix alignment calibrates retrieval operations that use prefixes. With fixed candidate budgets and ranking, original assignments on Beauty and Instruments expose more targets than controls preserving code capacity and coarse popularity. Resolving identical generated paths also shows that an explicit item selection rule recovers part of the gap between target path survival and singleton credit. Third, a released refresh repairs missing codes, while model adaptation separately restores new item recall. SIDScope supplies code, diagnostic summaries, and verification tools to compare mappings, evaluate their use in retrieval and generation, and test model reuse after changes.
Read the full abstract
Human-Computer Interaction
Engineering Psychological Safety in Autonomous Vehicles: A Systems-Theoretic Framework for Psychological Safety in Autonomous Vehicles and its Validation in Real-World Scenarios
Despite rapid technological advances, the societal acceptability of autonomous vehicles (AVs) remains limited by psychological barriers that extend beyond traditional concerns of physical safety. While factors such as trust and perceived safety are known to influence user acceptance, there is a lack of formalized mechanisms and engineering methods to systematically identify, assess, and mitigate psychological risks arising from human-AV interactions. To address this gap, this work proposes and validates a systems-theoretic framework for the assessment of psychological safety in autonomous vehicles. First, a comprehensive psychological safety risk model is defined, extending the Systems-Theoretic Accident Model and Processes (STAMP) to incorporate key psychological constructs such as trust, perceived control, predictability, and perceived support. Based on this model, a hazard analysis method (AV-PsySafe) is developed to systematically identify psychological hazards, unsafe control actions, and loss scenarios, while introducing a Psychological Safety Integrity Level (PsySIL) to support risk prioritization. Second, the applicability and relevance of the framework are evaluated through its deployment in realistic autonomous vehicle scenarios. A structured validation approach is implemented, including a methodological guide, standardized analysis templates, and the collection of analyst feedback. The results demonstrate that the framework can be consistently applied by practitioners, producing meaningful insights into psychological risks. Overall, this work establishes both the theoretical foundations and practical feasibility of a unified approach to co-assessing psychological and physical safety in autonomous systems, contributing to more human-centred and trustworthy AV development.
Read the full abstract
Machine Learning
GraphK: Variable-Size Graph Generation with Efficient Edge Construction
Graph generation models have advanced significantly with deep learning, yet they remain limited in scalability, flexibility, and ability to model underlying structures. We present GraphK, a novel encoder-sampler-decoder framework for graph generation that overcomes these challenges through structural flexibility and computational efficiency. Unlike autoregressive approaches constrained by vocabulary size (i.e. number of nodes in graph generation), GraphK allows for both upscaling (generating graphs with more nodes than the input) and downscaling, providing a flexible control over output graph size. By learning permutation-invariant latent representations and sampling new node embeddings via maximum likelihood estimation, GraphK generalizes across graph sizes and structures. For edge generation, we employ edge prediction with a KDTree-based top-k neighbor search in the latent space, reducing computational cost. Based on the manifold smoothness assumption, our method effectively captures graph properties. Experiments on synthetic and real-world datasets show that GraphK outperforms existing methods, accurately learns graph structures, and generates synthetic graphs without explicit definitions.
Read the full abstract
Computer Vision
MIFR: A Modality-Invariant and Fair Representation Framework for Skin Disease Classification
Skin diseases represent a major global public health burden, yet machine learning tools developed to assist in their diagnosis suffer from two critical limitations: reliance on only one modality for diagnosis and systematic performance disparities across skin tones. While existing approaches address each challenge separately, this work proposes a modality-invariant framework with fair representation (MIFR) for skin disease classification. The architecture pairs clinical photographs with dermoscopic images using ViT-based encoders, projecting each input into a high-dimensional embedding space via modality-specific projection heads. The resulting model is trained with a five-component multi-objective loss including weighted cross-entropy for classification, confusion and skin-type classification losses for fairness, per-modality supervised contrastive loss for class alignment, and a modality-invariance loss for clinical and dermoscopic modality alignment. Experiments on the HIBA+Derm7pt paired dataset and the external PAD-UFES-20 and ISIC 2019 datasets showed that modality-invariant representation learning provides competitive predictive performance compare to relevant baseline models and competitive fairness on the internal dataset. t-SNE visualizations confirmed that clinical and dermoscopic embeddings of the same disease are geometrically aligned, validating the joint objectives.
Read the full abstract
Data Structures and Algorithms
Decisive Margins in Differentially Private Voting
Differential privacy protects individual voting records by injecting randomness into the published outcome, but this noise can lead to erroneous results when an election is close. We study how precise central differential privacy and local differential privacy can be for common voting rules, including Plurality, Condorcet, Maximin, Plurality with Runoff, and Single Transferable Vote (STV). Our measure of precision is the margin of victory needed for a private mechanism to return the same winner as the non-private rule with high probability. We give private algorithms for publishing the winner and prove upper bounds on the required margin for these algorithms. We also prove lower bounds showing that nontrivial margins are necessary; many of these bounds match the corresponding upper bounds up to logarithmic factors. For STV, an information-theoretic upper bound matches the lower bound, but we prove that this guarantee cannot be achieved in polynomial time unless NP $\subseteq$ BPP. This gives a rare example of a computationally tractable task that becomes intractable when one simultaneously requires differential privacy and utility.
Read the full abstract
Machine Learning
To Go Far, Go Together: Diverse Preferences Induce a Curriculum for Reward Optimization
Learning a reward model from human feedback and optimizing a policy against it is one approach to aligning AI systems with individual users. From a fairness perspective, existing work improves such alignment by developing data-efficient and accurate reward models that capture minority preferences despite scarce data. We push this line of inquiry one step further and argue that data-efficient and accurate per-user reward models are not sufficient: users whose reward models are difficult to optimize at the policy level can become a new underserved group. We start from the observation that one user's reward model can be easy to optimize from the initial policy while another's is not. We argue that, given a sufficiently diverse user population, a curriculum naturally emerges between easy- and hard-to-optimize reward models. Building on this insight, we propose CurriPO, which grows a tree-structured curriculum to accommodate diverse user-specific objectives, covering the population in a single traversal. Specifically, CurriPO automatically constructs a curriculum over diverse user reward models, allowing it to branch from the existing curriculum and reuse reward models previously incorporated into the curriculum. To the best of our knowledge, this is the first work to explicitly exploit multi-user structure to address optimization in AI alignment. Extensive experiments on personalized continuous control in a simulated environment show that CurriPO achieves $1.2$–$2.1\times$ the population satisfaction of the strongest baseline while substantially reducing training time. Additional analysis attributes much of this improvement to the users left underserved by conventional optimization.
Read the full abstract
Discrete Mathematics · Cross-listed from math.OC
An Integer Programming Approach to Compute Lower Bounds for Ramsey Numbers Using Circulant Graphs
The Ramsey number $R(m,n)$ is the smallest order at which every red-blue edge coloring of a complete graph must contain a blue clique (a complete subgraph) of size $m$ or a red clique of size $n$. Determining these numbers exactly is extremely hard, and even certifying a lower bound requires exhibiting an explicit coloring that avoids both cliques. We develop an integer programming framework for certifying such lower bounds, restricting the search to circulant graphs, whose rotational symmetry lets us reformulate the problem in a projected distance space, reducing the number of binary variables from quadratic to linear in the graph order. We strengthen this projected model through coefficient reduction and solve it with a branch-and-cut algorithm whose separation routine exploits the common neighborhood structure of circulant graphs, combining heuristic and exact maximum-clique algorithms. In an extensive computational campaign on circulant graphs with up to 410 vertices, we improve the best lower bounds previously obtained by other methods by up to 11 points for 25 values of $R(3,n)$ with $24\le n\le49$ and $n\neq27$, each backed by an explicit graph certificate that can be independently verified with a stand-alone exact clique solver. To the best of our knowledge, our method also provides the first reproducible optimization-based procedure for certifying circulant Ramsey numbers $R_C(m,n)$, which we use to establish eight new values of $R_C(3,n)$ with $13\le n\le20$. Our framework, graph certificates, and stand-alone checker are provided as supplementary material to support independent verification and reuse.
Read the full abstract
Computation and Language
Readable, Faithful, Used: Three Dissociable Properties of Demographic Identity in a Language Model
Large language models are widely used to simulate survey respondents, yet their outputs are homogeneous and unfaithful to real inter-group differences, and whether this reflects what a model knows or uses has remained untested. Using representational similarity analysis against Pew American Trends Panel ground truth, we score demographic read-out locations in Mistral-7B and intervene causally across six attribute types. The internal geometry is faithful: attention-head read-outs dominate the standard residual read-out, reaching selection-corrected $\rho$ up to 0.63 – about 70% of the measurement-reliability ceiling – and one head, L11 H16, is significantly faithful across all six types, though race-based types stay weak and prompt-fragile, replicating in a second model family. Yet causal use does not track fidelity: the clearest causal pathway ($p=0.002$) sits in one of the least faithful types, the most faithful type shows no correction-surviving effect, and full identity swaps in the prompt move predictions by under 2% of their error. A 128-dimensional probe on that head lands 21-31% closer to survey truth than the model's answers, yet recovers almost none of the per-question group ordering. Readable, faithfully arranged, and causally used are three dissociable properties of the same model; treating them as one claim is what keeps the "can LLMs simulate populations" debate unresolved.
Read the full abstract
Computation and Language
Gradient Mirage: Trainable yet Label-Unidentifiable Gradients in Large Language Model Split Learning
Gradient matching attacks (GMAs) in LLM split learning (SL) rely on a critical yet underexplored assumption: the gradient exposed at the split interface is a faithful derivative of the client's full-label training objective. This gradient-objective consistency allows a curious server to recover private labels by searching for a sequence whose induced gradient explains the observation. We propose Gradient Mirage, a defense that breaks this consistency without discarding the optimization utility of the backward signal. Our key idea is to induce the adversary to solve a misspecified inverse problem, in which no plausible label sequence in the sequence space can explain the observed gradients. Concretely, Gradient Mirage achieves this by inducing inconsistency across three dimensions: objective, direction, and scale. Selective Autoregressive Supervision derives the exposed gradient from a masked surrogate loss rather than the full-label objective assumed by the attacker; Scale Blinding then applies randomized multiplicative rescaling, obscuring the gradient's natural magnitude; and Directional Privatization further randomizes the gradient direction while preserving its magnitude through the von Mises-Fisher (vMF) mechanism under a directional metric differential privacy guarantee. Crucially, utility is preserved: the Top segment still learns from all target tokens via Dual-Track Backpropagation, the exposed gradient remains informative since each supervised token retains its complete autoregressive context, and Bottom-Gradient Recovery restores the effective gradient for Bottom-segment optimization. Extensive experiments show that Gradient Mirage provides substantially stronger protection than existing defenses under comparable fine-tuning performance, achieving a better privacy-utility trade-off.
Read the full abstract
Numerical Analysis
A shifted energy barrier approach for phase-field modeling of tensile-dominated brittle fracture
The classical AT1 phase-field model contains an intrinsic energy barrier for crack nucle ation, which makes the predicted strength depend on the fracture toughness and the regularization length. For tensile-dominated brittle fracture, this barrier is shifted by mapping the Rankine criterion, evaluated on the effective stress, onto a state-dependent active-energy threshold. The prescribed tensile strength then controls crack nucleation, while the AT1 crack-density functional, stiffness degradation, and degraded stress response remain unchanged. Since the threshold depends on the current stress state, the field equations are derived from a restricted variational principle. A microforce formulation identifies the barrier shift as a dissipative resistance and provides the corresponding lower bound on the regularization length. In one-dimensional tension, closed-form solutions recover the prescribed peak strength and give a cosine-type localization profile that ap proaches the classical AT1 profile as the shift vanishes. Numerical examples show that, within the admissible range, the nucleation load is nearly insensitive to the regularization length and the predicted multiaxial nucleation states follow the Rankine envelope. Under overall compression, crack nucleation remains associated with local tensile stress concen trations. The formulation also captures the transition from strength-controlled failure for small flaws to the LEFM limit for large cracks.
Read the full abstract
Computation and Language
Learning Canonical Register Automata over Ordered Data Domains
Register automata are finite automata equipped with memory that recognize data languages over infinite alphabets. In this work, we investigate active learning algorithms for deterministic register automata (DRAs) over ordered data domains–covering both dense domains, such as the rationals, and non-dense domains such as the integers. We show that the active learning problem for DRAs over both dense and non-dense ordered domains can be treated within a single unified framework. More specifically, we develop and implement a polynomial-time active learning procedure for DRAs over ordered domains, using oracles for membership, equivalence and memorability queries. The memorability queries were originally introduced for learning DRAs over domains with identity tests. Our unified framework also leads to a new consequence: minimization of DRAs over the non-dense ordered domain of integers is decidable, extending a result previously known only for dense domains. Finally, we give improved complexity bounds of several decision problems for DRAs over ordered domains that are closely related to the queries used in active learning.
Read the full abstract
Information Retrieval
GateDiffInt: Gate-Mediated Controllable Diffusion and Multi-Intent LLM Distillation for User Behavior Modeling
Existing ranking models encode intent only implicitly, making it hard to disentangle structured intents of varying strength and temporal scale. Noise and intent in behavior sequences are mutually reinforcing—we call this Noise–Intent Coupling (NIC). Noise dilutes true intents, while the lack of structured intent priors leaves denoising without a clear target. To address NIC, we propose GateDiffInt, an intent interaction framework for industrial ranking. It uses the final conversion signal to jointly align sequence denoising and intent extraction. GateDiffInt applies a controllable forward diffusion process with dual gating to enhance and denoise behavior sequences. A large language model then acts as teacher to distill four structured intents—long-term, short-term, latent, and conversion into a lightweight student model. The enhanced sequence and structured intent representations are deeply fused via attention to produce intent-aware representations for conversion-rate prediction. Extensive experiments on public and large-scale industrial datasets show consistent gains over strong baselines. In online A/B tests serving hundreds of millions of daily active users, GateDiffInt delivers substantial GMV improvements and has been deployed to primary traffic, confirming both effectiveness and production readiness.
Read the full abstract
Machine Learning
Enhancing Distance-Based Graph Autoencoders with Structural Penalties for Dynamic Graph Embedding
Graph autoencoders (GAEs) are widely used for learning representations of dynamic graphs. However, their optimisation objectives typically do not take structural heterogeneity across nodes into account. We propose three distance-based GAE variants that incorporate structural penalties into the reconstruction loss. All variants share a two-layer Graph Convolutional Network encoder and a Euclidean-distance decoder trained with distance-based reconstruction objectives. We extend sparsity-corrected loss with two node-level regularization terms: (i) a hub penalty based on degree centrality, and (ii) a penalty based on Natural Community Local Intrinsic Dimensionality (NC-LID). The paper is motivated by prior evidence linking high NC-LID to reduced embedding quality. The proposed methods are designed to emphasize reconstruction errors for structurally ambiguous nodes. Experiments on multiple dynamic graph data sets show that incorporating NC-LID-based regularization consistently improves reconstruction performance over the baseline without structural regularization and the method using hub-aware regularization. These findings highlight NC-LID as a useful structural signal for enhancing distance-based graph autoencoders in dynamic settings.
Read the full abstract
Machine Learning
Beyond Predictive Fairness: Quantifying Attribution Consistency Across Demographic Groups in Diabetic Retinopathy Screening
Fairness in medical imaging is commonly evaluated through subgroup performance metrics, yet it remains unclear whether models rely on consistent visual evidence across demographic groups. This work introduces the Explanation Consistency Score (ECS), a fairness-aware metric based on Jensen-Shannon divergence that quantifies the similarity of attribution maps across subgroups. Using diabetic retinopathy screening as a case study, ECS is evaluated globally and within disease severity. Experiments reveal that while predictive performance differs across ethnic groups, explanation consistency remains relatively high and shows no significant association with performance disparities. These findings suggest that predictive fairness and explanation consistency capture complementary dimensions of model behavior, motivating fairness evaluations that extend beyond predictive performance.
Read the full abstract
Computers and Society
Epistemic Subordination: Generative AI and the Infrastructure of Knowledge
Generative AI does not merely produce biased outputs. It encodes the majority's way of knowing as the default infrastructure of knowledge itself. We call this epistemic subordination. The training process compresses the full breadth of human expression into a single probabilistic model whose statistical baseline reflects the languages, assumptions, and cultural frameworks of the dominant culture. Minority epistemologies are not excluded but absorbed: present in the training data, yet structurally subordinated in the output. The result is not a collection of discrete biases that can be audited and corrected. It is an epistemic condition embedded in the architecture from which all outputs emerge. This unified harm cuts across three legal domains – anti-discrimination law, cultural and linguistic rights, and democratic viewpoint pluralism – and each fails to address it for the same structural reason: existing law regulates downstream, at the level of decisions and applications. The remedy must match the site of harm. If epistemic subordination is produced at the level of model training, then law must learn to govern at that level.