Software Engineering
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Monday 31 August
The software engineering research community has enthusiastically embraced the integration of Large Language Models (LLMs) into complex techniques to solve a wide variety of tasks. However, the extent to which this investment is strategic remains unclear, as the native capabilities of successive frontier model generations can rapidly render existing techniques obsolete. To assess this research investment, we analyze 35 LLM-based technique papers from ICSE 2026. We evaluate whether their complex tools can be outperformed by the simplest possible alternative: a single, automatically generated prompt executed on a newer generation model, without any iterative refinement. We find that for between 37% and 63% papers, a newer model with a single prompt natively outperforms the heavily engineered tooling proposed just a year prior. We identify that constructive techniques like code generation or repair are more amenable to substitution by a single-prompt. We also identify a surviving set of papers relying on strategies that provide additional insights to the model where newer LLMs will amplify the proposed technique. Our findings raise questions about the cost-benefit proposition of techniques designed as workarounds to temporary model deficits and the need to focus on enduring challenges that scale synergistically with future model generations. Our source codes and results are made publicly available at https://github.com/less-lab-uva/What-Survives-the-Next-Model.
Superconductivity
Comment on "Distinct Behaviors of Inner and Outer CuO$_2$ Planes in Quadruple-Layer Cuprate (Cu,C)Ba$_2$Ca$_3$Cu$_4$O$_{11 + \delta}$"
In a recent Letter, Sun et al. report photoemission spectroscopy measurements on the four-layer cuprate (Cu,C)Ba2Ca3Cu4O11 (CuC-1234) in which they resolve two superconducting gaps associated with the inner (IP) and outer (OP) CuO2 planes. CuC-1234 has a Tc as 117 K in a sample previously structurally characterized using neutron powder diffraction. From the Luttinger sum, the beta-band of the IP was found to be heavily underdoped (p~0.07) while the OP alpha2 bonding band was strongly overdoped (p~0.25). Each revealed a gap with very different momentum and temperature dependences. A large gap on the beta band was found to close at the bulk Tc, while a smaller gap on the alpha2 band closed at 70 K. Such two-gap behavior would indicate weak-coupling between the IP and OP, resulting in a second gap-opening temperature, Tc2, and implying a two-step development of the superfluid density on cooling, first on the IPs then on the OPs. Here, we point out that the observed values of Tc and Tc2 are fully consistent with a long-standing correlation of Tc with a bond-valence sum parameter, V+, calculated from the crystallographic bond lengths. This suggests that Tc2 in the OP is entirely consistent with the local structure of the OP where the value of V+(OP) is reduced by the presence of the apical oxygens and Tc2 reduced accordingly. It is as though two largely decoupled superconductors are present in the single 1234 compound and their individual pairing temperatures are determined by their local nearest-neighbour structure, reflecting the very short-range physics.
Read the full abstract
Numerical Analysis
Shape Holomorphy and Sparse Approximation of the Maxwell Electric Field Integral Operator
Time-harmonic Maxwell scattering by a perfectly conducting obstacle is governed by the electric field integral equation (EFIE), whose natural energy space is $\boldsymbol H^{-1/2}(\mathrm{div}_\Gamma,\Gamma)$. For countably parametrized surface deformations, both the operator and its energy space depend on the geometry. We establish operator-valued shape holomorphy of the EFIE operator on a fixed reference space and derive dimension-independent sparse approximation rates. The main analytical difficulty is a fractional mapping property absent from existing $L^2$-based operator-valued shape-holomorphy theory for weakly singular kernels: the Maxwell graph norm requires uniform holomorphy of the complex-deformed scalar single layer as an operator from $H^{-1/2}$ to $H^{1/2}$. We prove this one-order smoothing by realizing the Laplace principal part as the trace of a uniformly sectorial complex-coefficient divergence-form problem on a fixed real ambient domain. Combined with a surface contravariant Piola transformation, which identifies the geometry-dependent Maxwell trace spaces and cancels the surface Jacobians in the pulled-back EFIE exactly, this yields $(\mathbf b,p,\varepsilon)$-holomorphy of the EFIE operator family for $\ell^p$-summable shape deformations, $0<p<1$. Under pointwise exclusion of interior electric resonances, parameter-uniform invertibility is derived rather than assumed, and the surface current and far field inherit the same parametric regularity. Their Legendre coefficients are $\ell^p$-summable, yielding dimension-independent best $N$-term approximation rates. The operator-level results also transfer to Galerkin boundary element operators on a single reference mesh, with constants independent of the discretization dimension, and provide reusable surrogates for multiple incident fields and bounded linear observables.
Read the full abstract
Signal Processing
A Deep Iterative Refinement Receiver for OTFS Symbol Detection in Doubly-Dispersive Channels
Orthogonal time frequency space (OTFS) modulation has emerged as a promising candidate for high-mobility wireless communication systems due to the diversity it offers across both time and frequency. Reliable OTFS detection, however, remains challenging under doubly-dispersive channels, where delay and Doppler dispersion induce structured interference between transmitted symbols and complicate symbol recovery. To address these challenges, we propose a two-stage iterative OTFS detector that integrates a physics-informed learned initializer with an iterative refinement network, enabling progressively more accurate symbol estimates in doubly-dispersive channels. The initializer incorporates the known delay-Doppler input-output relationship to produce a robust first-stage estimate, while the refinement stage iteratively suppresses residual symbol interference in the delay-Doppler domain. Simulation results demonstrate that the proposed detector achieves consistent performance gains over conventional and existing learning-based detectors across a variety of channel conditions. These results highlight the effectiveness of incorporating known channel structure into the detection process and using iterative refinement for improved and robust OTFS detection.
Read the full abstract
Cryptography and Security
Does Reasoning Mitigate Backdoor Attacks? A Neuro-Symbolic Perspective
Neuro-Symbolic (NeSy) AI has recently emerged as a novel paradigm to enable trustworthy AI, aiming at integrating sub-symbolic neural perception with grounded symbolic reasoning. The neuro-symbolic integration process that characterizes these models has been proven beneficial to achieve more transparent, explainable and efficient AI systems. Meanwhile, their properties under adversarial settings have been overlooked being frequently deemed robust-by-design. However, the neural-symbolic integration process they leverage constitutes an additional layer of complexity that may provide an attack entry-point. Therefore, in this paper, we claim that an in-depth investigation of the adversarial robustness of NeSy models is necessary and provide the first systematic evaluation of backdoor attacks against NeSy. To this end, we compare the most popular NeSy framework, namely DeepProbLog, against baseline neural networks across a total of eight backdoor settings and four reasoning tasks. Our experimental results show that while NeSy models are indeed more robust than their neural counterpart on average, their robustness vastly depend on the strictness of the reasoning process being enforced and its compatibility with the chosen adversarial target. The source code to reproduce our experiments is made available at https://github.com/marcoantoniocorallo/NeSy-Backdoor.
Read the full abstract
Computation and Language
Toppling the Hierarchy in Byte-level Language Modeling
This work examines recent byte-level models and their failure to perfectly manipulate characters. State-of-the-art byte-level models use a hierarchical structure, starting at the byte level, downsampling to the word level, and then upsampling back to bytes. While this improves training and inference efficiency, we find that the hierarchical design itself limits character-level understanding, with pure byte-level models consistently outperforming hierarchical variants on character manipulation tasks. Ablating transformer layers into attention and feed-forward components further reveals that byte-level attention is the primary mechanism driving this behavior. Together, our results provide an explanation for the character-level failures of hierarchical byte models and establish a clear trade-off between computational efficiency and fine-grained character understanding.
Read the full abstract
Operator Algebras
Two-basepoint Terwilliger algebras and the quantum symmetry of prime-order circulants
Which vertex-transitive graphs of prime order have quantum symmetry? The question of Banica, Bichon and Chenevier is open in the dense regime of Paley graphs, where coherent-algebra methods give no information. To each such graph we attach a two-basepoint Terwilliger algebra of its cyclotomic scheme and study the module it generates from the basepoints: fullness forces the quantum permutation algebra to be commutative, and the module admits no intermediate state, containing either exactly two point masses or all $p$ of them. One point mass, captured at any depth, therefore suffices, and Chassaniol's orbital criterion is the depth-one case. Three consequences follow. A sharp counting argument replaces the Banica–Bichon–Chenevier threshold $p>6^{\varphi(k)}$ by the quadratic bound $p>(k-1)(k-2)+2$, where $k$ is the type. Four certificates, each a short list of additions modulo $p$, settle $C_{31}(2,4,8,15)$ and $C_{41}(4,10,16,18)$, the two graphs left open by Chassaniol, and complete the classification for type at most $10$ without machine assistance. An exact computation extends the dichotomy “quantum symmetry if and only if complete or empty” to all prime orders $p\le250$, settling the Paley graphs $P_{p}$ with $p\le241$, the first beyond $P_{17}$. What remains is the capture of a single explicit vector: the midpoint $2^{-1}$ of the two basepoints.
Read the full abstract
Fluid Dynamics
Deforming Vortex Force Mapping for Undulating Swimmers
Undulatory propulsion is commonly interpreted through reactive added-mass loading associated with body kinematics and vortex-induced loading, but their relative contributions to the streamwise force have not been systematically quantified across swimming kinematics. Moreover, existing vortex-based descriptions do not resolve the signed spatial contribution of vortical flow to the streamwise force. We extend vortex force map (VFM) attribution to deforming bodies, demonstrated for prescribed two-dimensional undulating swimmers at $Re=5000$ across a range of Strouhal numbers and dimensionless wavelengths. The deforming-body VFM organises the streamwise force as a positive reactive added-mass scale with a signed vortex-pressure modification. Across the 80 cases, vortex pressure opposes the reactive contribution in 76 cases. Reference-configuration maps show that the near-body vortex-pressure contribution is consistently adverse, whereas the wake contribution changes sign across regimes and becomes favourable at high thrust. Net vortex-pressure assistance occurs only for the high-thrust anguilliform swimmer. A matched high-thrust pair further shows that the case with the higher peak downstream velocity has the smaller signed wake contribution, demonstrating that wake intensity alone does not determine its force effect.
Read the full abstract
Machine Learning
Context Window Failures in Relational Foundation Models
Recent Relational Deep Learning architectures have been proposed as foundation models for multi-table relational data, yet they impose constrained neighborhood budgets that force row truncation when an entity has many related records. We introduce Animus, a synthetic financial dataset in which predicting customer income requires aggregating up to tens of thousands of transactions. On the raw representation, three recently proposed models (RT, Griffin, RelGT) achieve $R^2 \le 0.18$; a single, routine, temporal pre-aggregation step recovers $R^2$ up to $0.65$. This questions whether current relational foundation models are ready for high-cardinality real-world data.
Read the full abstract
Solar and Stellar Astrophysics
Age Discrepancy in Three Galactic Cepheid Binaries
SV Per, RW Cam, and KN Cen are Galactic classical Cepheids found in binary systems along with B-type companions. Each system exhibits an “apparent age discrepancy", where the Cepheid appears significantly younger than its companion. We compute stellar evolution tracks using the MESA code and compare them with these stars on the Hertzsprung-Russell diagram to estimate the masses and the corresponding ages. We find that, for two of the systems, the companions appear to be between 1.5 times to double the age of the Cepheids, suggesting that the Cepheids may be rejuvenated merger products of two main-sequence stars that have since evolved. We find that the third system KN Cen is consistent with single star evolution when we reevaluate ultraviolet observations and find that the companion is hotter than previously measured. We test this idea using the MESA code for each system by evolving a single star model and rapidly accreting mass on the main sequence to emulate a merger, and then continuing the evolution until the model is consistent with the parameters of the Cepheids. The results are consistent with the hypothesis that the Cepheids were originally merger products. We conclude with a discussion of potential observational tests of the merger hypothesis.
Read the full abstract
Earth and Planetary Astrophysics
An Upper Limit on Turbulent Viscosity in the Circumjovian Nebula
Circumplanetary disks are open, dynamic systems sustained by a meridional circulation that draws gas and dust from the parent nebula along nearly polar streamlines, and ultimately returns some of this material through a viscously driven, in-plane outflow. In such a disk, the water ice line acts as a site of solid-mass accumulation: inward-drifting icy solids sublimate interior to the ice line, the resulting vapor is advected outward by the gas, and recondensation beyond the ice line resets the material to a small Stokes number. The operation of this drift-mediated loop requires the fragmentation-limited Stokes number of the mass-dominant icy aggregates to exceed the equilibrium Stokes number at which the radial drift of solids reverses. Because the former scales inversely with the Shakura-Sunyaev viscosity parameter $\alpha$ while the latter is directly proportional to it, this requirement yields a compact upper bound on the vigor of turbulence in the satellite-forming region. Adopting an actively heated, steady-state circumjovian disk truncated at the tidal radius, we find a marginal bound of $\alpha\lesssim 10^{-3}$ for an icy-particle fragmentation threshold of $v_f\simeq1\,{\rm m\,s^{-1}}$ - appropriate for cold ice with sticking properties similar to silicate dust. This bound scales linearly with $v_f$, reaching $\alpha\lesssim5\times10^{-3}$ for more adhesive, warm ice ($v_f\simeq5\,{\rm m\,s^{-1}}$). The bound carries no explicit dependence on the disk's mass flux, and is set entirely by the local thermal state, pressure gradient, and geometry of the circumjovian nebula.
Read the full abstract
Machine Learning
Can LLMs Use Relational Transformer Embeddings?
Injecting frozen relational-encoder embeddings as soft tokens into a large language model (LLM) is a conceptually appealing fusion strategy: the encoder handles multi-table structure, the LLM handles language and reasoning, and no lossy text serialization is required. We test this hypothesis concretely by injecting embeddings from a frozen Relational Transformer (RT) into Qwen3.5-4B via a learned MLP projection and LoRA adaptation, trained first with supervised fine-tuning (SFT) on chain-of-thought reasoning traces and then with group-based reinforcement learning (GSPO). We evaluate across 10 binary classification tasks on 6 relational databases from RelBench, under four supervision regimes: single-task (ST), within-dataset (WD), cross-dataset (CD), and all-task (ALL). The hybrid model does not consistently outperform standalone RT: it is frequently below random, highly sensitive to serialization format and relational-token budget, and unstable under RL training. We report these negative results and analyze the failure modes, arguing that soft-token fusion requires stronger alignment objectives and schema-aware design before it can serve as a reliable route to relational prediction.
Read the full abstract
Statistical Methodology
Genetic association testing with multivariate survival phenotypes under interval censoring
Set-based genetic association tests provide a powerful framework for detecting genetic effects on complex traits by jointly analyzing multiple genetic variants. Although set-based methods have been developed for interval-censored survival outcomes, existing approaches primarily focus on a single survival phenotype and therefore do not fully use information from multiple correlated outcomes. In this paper, we develop two Weighted V Tests for Multivariate Interval-Censored Data (WV-M-IC), extending the weighted V-statistic framework (Wu et al., 2021) to the joint analysis of multiple correlated interval-censored survival outcomes. The performance of these methods is evaluated through simulation studies, showing that the proposed approaches can provide power gains compared with single-outcome analyses. We apply the proposed methods to the ZOE 2.0 study to investigate dental caries progression in children.
Read the full abstract
Artificial Intelligence
Towards a Belief-Based World Model for LLM Agents
Large language models (LLMs) are being used as policies for autonomous decision-making and planning in many domains. Despite their strong reasoning capabilities, LLMs struggle with long-horizon tasks, especially under partial observability. World models are a promising way to enhance policy performance, both during training and inference. During inference, agents currently use world models to simulate the consequences of candidate actions before choosing an action, which can improve decision-making. However, we argue that simulation alone is an incomplete interface for decision-making under partial observability: simulation does not adequately capture uncertainty about the current state, which agents may need for accurate decision-making. We address this limitation with Belief-Based World Models (BB-WMs), which maintain a belief that LLMs can query to access information on what is known and uncertain about the current state. Before developing methods to learn accurate BB-WMs, this paper focuses on a more fundamental question: does exposing a world model's belief directly to an LLM policy improve decision-making? Our results show that giving LLM agents access to beliefs improves task performance under partial observability, while remaining complementary to existing simulation-based world models. Code: https://github.com/skumar-ml/belief-world-models.
Read the full abstract
Computation and Language
Location-Aware Language Models via Secondary Embeddings
Pretrained transformer-based language models achieve strong performance across a wide range of NLP tasks but remain limited in encoding geo-locational semantics, leading to suboptimal representations of place names and spatial entities. In this work, we propose a lightweight, model-agnostic approach for injecting geo-spatial awareness into pretrained embeddings without modifying the tokenizer or requiring costly retraining. Our method augments input representations with structured geographic signals by combining location names with their corresponding latitude and longitude, and employs a location-focused masking to better align textual representations with real-world spatial relationships. This design allows the model to incorporate geo-spatial context while preserving existing semantic and syntactic knowledge. Experimental results demonstrate substantial improvements in geo-spatial alignment while maintaining comparable performance on standard NLP benchmarks such as GLUE. The method is computationally efficient, requiring only minutes of additional training, and generalizes across multiple model architectures and scales.
Read the full abstract
Artificial Intelligence
mimeo: Compiling Public Expert Corpora into Agent Skills and Testing What Transfers
Giving an agent a file about a named expert can supply hard-to-find material, produce a recognizable persona, or change what the agent decides. These are different claims. We test each one. mimeo is an open-source tool that finds a person's public work, checks each extracted quotation against the cached source text, and writes a file an agent can load. Eight logged builds averaged 38 model calls; the check rejects 13.2% of extracted quotations. We tested four expert files with one coding-agent harness. Knowledge access was clearest: mimeo answered all 20 obscure, quotation-heavy questions; no closed-book condition answered more than 10. Keyword search (BM25) over the same pages answered 15-17, a gap this sample cannot resolve. Grounding showed one clear benefit: personas written from model memory misstated a documented position on 1-4 of 20 answers under every grader; the plain agent and mimeo never did. Every persona was easy to spot on short open prompts, and adding task material lowered identification by 18-23 points. mimeo was no more identifiable than a from-memory profile. Judgment transfer remained unresolved because both tests hit their ceiling: every condition found 94-97% of the problems planted in engineering tasks and scored 94-100% on 16 new application scenarios. An AI-judged "sounds like the expert" score changed with the judge: two of four preferred answers based on a model's stereotype, while two found no difference on the same text. That is a caution against relying on a single AI judge. The evidence supports mimeo as a compact, inspectable reference on a person, not as a demonstrated transfer of their judgment. Toolkit and expert profiles: https://github.com/K-Dense-AI/mimeo
Read the full abstract
Operator Algebras
Rapid decay and functional calculus in ${\rm C}^*$-probability spaces
Let $(A,\rho)$ be a tracial ${\rm C}^*$-probability space with the rapid decay property relative to a filtration $L$. We show that the associated Sobolev algebra $H_L^{\infty}(A,\rho)$ is closed under the smooth functional calculus of $A$. Some consequences include norm estimates, the ideal separation property, and the fact that the inclusion $H_L^{\infty}(A,\rho)\subset A$ induces an isomorphism in $K$-theory.
Read the full abstract
Dynamical Systems
Pushforward dynamics on Wasserstein spaces and measure rigidity
For an endomorphism $\phi$ of a closed Riemannian manifold $M$, we study the pushforward action $\phi_\ast$ on the Wasserstein space $\mathcal{P}(M)$ at a measure $\mu_0$ preserved by $\phi$. We show that if $\phi$ is a $C^2$ covering map and $\mu_0$ has positive $C^1$ density, then $\phi_\ast$ is Gâteaux differentiable at $\mu_0$ along tangent directions, with derivative given by the transfer operator of $\phi$ acting on vector fields, followed by orthogonal projection onto the tangent space. The derivative is the adjoint of the Koopman operator restricted to the tangent space, and its fixed space consists of the directions in which $\mu_0$ can be deformed while preserving invariance to first order. For appropriate pairs of endomorphisms of $\mathbb{T}^d$, we compute the intersection of their fixed spaces, show that it contains an infinite family of linearly independent continuous vector fields, and construct, for every $n$, an embedded $n$-dimensional family of measures that are nearly invariant under both endomorphisms. First-order rigidity therefore fails in every dimension. In particular, rigidity phenomena such as higher-dimensional analogues of the Furstenberg conjecture, if true, are genuinely nonlinear.
Read the full abstract
Machine Learning
HBQ: Hierarchical Scaling Block Quantization with Hardware-Efficiency-Aware Design for Accurate LLM Inference
Block Quantization (BQ) enables efficient LLM inference by quantizing both weights and activations, but its design space remains underexplored. Through hardware-accuracy design space exploration, we identify block size as a key trade-off: larger blocks improve hardware efficiency by amortizing dequantization and accumulation costs, but degrade accuracy. Motivated by this insight, we propose Hierarchical Block Quantization (HBQ), which combines large blocks with low-overhead significand (SIG) scaling for second-level quantization. SIG scaling effectively compensates for large-block quantization errors while accounting for distinct weight and activation distributions. HBQ-A achieves W4A16-level accuracy with W4A5 and lower area than NVFP4, while HBQ-E further reduces hardware cost by 17% while outperforming existing BQ methods in accuracy. We implement HBQ for weights, activations, and KV cache in a 28nm ASIC accelerator and introduce partial-sum BQ to reduce EMA energy. At comparable accuracy, HBQ achieves 2.3x/4.6x higher area/energy efficiency than state-of-the-art weight-only quantization and 1.6-3.3x lower system energy with 1.5-3x speedup over prior BQ methods. Our implementation is publicly available at: https://github.com/SeoLabCornell/HBQ.git.
Read the full abstract
Neural and Evolutionary Computing
Investigating Hyperparameter Optimization and Transferability for ES-HyperNEAT: A TPE Approach
Neuroevolution of Augmenting Topologies (NEAT) and its advanced version, Evolvable-Substrate HyperNEAT (ES-HyperNEAT), have shown great potential in developing neural networks. However, their effectiveness heavily depends on the selection of hyperparameters. This study investigates the optimization of ES-HyperNEAT hyperparameters using the Tree-structured Parzen Estimator (TPE) on the MNIST classification task, exploring a search space of over 3 billion potential combinations. TPE effectively navigates this vast space, significantly outperforming random search in terms of mean, median, and best accuracy. During the validation process, the best hyperparameter configuration found by TPE achieves an accuracy of 29.00% on MNIST, surpassing previous studies while using a smaller population size and fewer generations. The transferability of the optimized hyperparameters is explored in logic operations and Fashion-MNIST tasks, revealing successful transfer to the more complex Fashion-MNIST problem but limited to simpler logic operations. This study emphasizes a method to unlock the full potential of neuroevolutionary algorithms and provides insights into the hyperparameters' transferability across tasks of varying complexity.
Read the full abstract
Algebraic Topology
Equivariant formality and representation theory
Let $G$ be a compact connected Lie group and $K$ its connected Lie subgroup. Using the $K$-theoretic version of equivariant formality developed in [F2] which involves analysis of vector bundles over $G/K$, we find two characterizations of equivariant formality of the isotropy action of $K$ on $G/K$. The first one equates equivariant formality with a "smoothness" condition of the restriction map of the representation ring of $G$ to that of $K$. The second characterization asserts that equivariant formality amounts to the equivariant $K$-theoretic index pairing of two certain special $K$-theory classes being nontrivial in some sense. By applying these characterizations, we are able to give a representation theoretic criterion for equivariant formality of the isotropy action by a circle subgroup, as well as an invariant theory criterion for the case where $K$ is a torus of dimension one less than the rank of $G$, culminating in a complete classification of equivariant formality where $G$ is further assumed to be simple.
Read the full abstract
Computer Vision
Instance-Guided Report Anchoring for Text-Free 3D Abnormality Segmentation in Chest CT
Accurate 3D abnormality segmentation in chest CT requires dense spatial supervision, but obtaining expert voxel-level labels is costly. Radiology reports, however, are routinely generated during clinical interpretation and contain instance-specific descriptions that can provide additional guidance without new dense annotation. Existing vision-language grounding methods typically require report-derived findings at inference, making localization dependent on paired text and limiting each forward pass to a queried finding. We propose Instance-Guided Report Anchoring (IGRA), a model-agnostic module that preserves the correspondence between each annotated abnormality instance and the report finding that describes it. IGRA pools each instance representation and anchors it to the corresponding finding embedding during training; all text-related components are discarded at inference. We further reformulate free-text grounding on ReXGroundingCT as multi-label volumetric segmentation by merging same-category instances, allowing all abnormality categories to be predicted in one image-only forward pass. IGRA improves Dice by 22.5% over the strongest image-only baseline (30.93 vs. 25.25) and is comparable to VoxTell on the single-finding subset (30.29 vs. 30.43). Applied unchanged to four standard 3D segmentation backbones, IGRA improves Dice and hit rate across all architectures. Zero-shot evaluation on LIDC-IDRI, PleThora, and a private in-house dataset further shows consistent gains over image-only baselines.
Read the full abstract
Machine Learning
CRAD: Class-wise Reliability-Aware Distillation for Decentralized Heterogeneous Federated Learning
Conventional federated learning (FL) relies on parameter averaging, which forces clients to be doubly homogeneous: it demands an identical architecture and degrades under non-IID data. Real-world deployments usually break both assumptions. We sidestep both by building a decentralized knowledge distillation framework in which each client evaluates its peers' model snapshots on its own local data and distills from the resulting soft predictions. Because knowledge is transferred through the shared class posterior, clients are free to run different architectures; and because every teacher is evaluated on the student's own device, raw data never leaves the client, with no central server or public dataset required. Within this setting, we identify and address an under-examined problem: how to combine the peer teacher predictions. Existing methods, like uniform averaging, ignore how knowledge reliability varies across teachers and classes. We propose Class-wise Reliability-Aware Distillation (CRAD), which, per class, first discards teachers that disagree with the peer consensus and then takes a weighted average of the rest, weighting each teacher by its per-class reliability (precision, or inverse variance). Since the variance of an accuracy from $n$ samples scales as $1/n$, support enters automatically: among the teachers that survive filtering, a teacher is trusted for a class to the degree that it is both accurate and well-evidenced for it. On CIFAR-10, CIFAR-100, and PathMNIST, across heterogeneous architectures under severe non-IID skew, CRAD consistently outperforms competing methods in global accuracy. Our code is publicly available at https://github.com/bbilbeis/CRAD.
Read the full abstract
Cryptography and Security
Capability-Gated Language Models: Security Composes, Utility Does Not
Deployed language model safeguards (safety fine-tuning, filtering, unlearning) vary by principal only outside the model weights: filters are reconfigured, tiers are multiplied, and artefacts are reissued; inside one set of weights every request meets the same model configuration. This motivates us to define capability-gated deployment: per-principal access control inside one set of weights, whose configurations form a lattice - meets accumulate a principal's restrictions and joins pool a coalition's reach. We instantiate it by sparse rank gating over an existing nested-factorisation mechanism, guide profile search with one-pass attribution, and read every result once from a pre-registered held-out split. Security approximately composes: provably exactly at meets under a monotone-elicitation assumption we falsify pointwise. In two lineages the median held-out meet deepens suppression; the one effect surviving correction strengthens it. Utility does not: individually harmless profiles can compose to retention and fluency damage, and no compositional bound exists.
Read the full abstract
Machine Learning
Group Adaptive Clipping Policy Optimization
Group relative policy optimization for reinforcement learning with verifiable rewards (RLVR) typically uses a fixed importance-sampling (IS) ratio clipping boundary across all rollouts. We identify a key limitation: rare correct rollouts on harder problems and abundant correct rollouts on easier problems are clipped at comparable rates, despite contributing very different learning signals. Rollouts with low group success exhibit larger IS ratios and carry stronger gradient signal for exploration and solving new problems, yet are disproportionately suppressed by fixed clipping. To address this, we propose Group Adaptive Clipping Policy Optimization (GAPO), a plug-in modification to GRPO methods that adapts the clipping boundary to the rollout advantage. GAPO is motivated by a reverse-KL trust-region perspective, which suggests that rollouts with larger learning signal should receive proportionally greater update headroom. GAPO requires no reward shaping and preserves the standard PPO/GSPO surrogate while adapting only the clipping threshold. Across Qwen and Llama models, GAPO consistently improves both Pass@1 and Pass@k over fixed clipping and advantage-shaping baselines on math reasoning and coding benchmarks where the pass rates by the base model are relatively low.
Read the full abstract
Computation and Language
(V)LMs generalize beyond surface co-occurrence: Evidence from cross-modal number agreement
Language models learn about grammatical number primarily from co-occurrence, and show frequency effects as a result—sometimes taken to indicate that they do not learn abstract “rules”, and are instead dependent on specific lexical items. Testing generalization with text stimuli alone cannot settle this debate, since distributional cues (is/are, this/these) easily give number away. We instead use cross-modal generalization as a tool to investigate abstractions in LMs that can also accept visual inputs (VLMs), restricting the evidence that diagnoses number to an extra-linguistic modality. We teach VLMs pairs of new nouns by adding new embeddings and only updating them during learning, comparing conditions where number is diagnosed by visual cues alone against ones where it is disambiguated by text. Across behavior, representational dynamics, and causal mechanisms, we find non-trivial evidence for cross-modal generalization across both exposure conditions, and that linguistic vs. extra-linguistic cue conditions are treated in similar ways in the internal mechanisms of the model. This suggests that statistical learners like VLMs can generalize beyond surface-level co-occurrence and show genuine abstraction-compatible behavior.
Read the full abstract
Distributed Computing
DRLM: Deep Reinforcement Learning-Based LLM Query Orchestration in Edge Environments
Large language model (LLM) services increasingly process heterogeneous queries with diverse latency, accuracy, and resource requirements. While edge deployment reduces response time, the heterogeneity of devices and the diversity of model families, parameter scales, and quantization levels make efficient LLM query orchestration challenging. This paper introduces DRLM, a Deep Reinforcement Learning-based LLM query orchestration framework in edge environments. DRLM integrates two lightweight predictors: (i) a class-conditioned quality estimator that maps queries to semantic categories and infers model performance, and (ii) a feature-driven latency predictor that estimates inference time across model-device configurations. These predictions, combined with system state, feed a factorized Proximal Policy Optimization (PPO) agent that performs state-aware orchestration decisions. To enable data-driven orchestration, we construct a large-scale benchmarking dataset with 223 835 measurements spanning 1258 queries, 6 query classes, 8 model families (32 deployed instances), 5 quantization levels, and heterogeneous edge devices. Evaluation on a 64-node edge cluster and comparison with three baselines and two state-of-the-art methods show that DRLM reduces inference latency by up to 51% and queuing delay by up to 67 %, while incurring at most 8% accuracy loss. It improves latency under increasing workloads up to 61.4%, demonstrating robust and stable orchestration.
Read the full abstract
Artificial Intelligence
Conversation Coach: A Voice-enabled AI System that Helps Practice Difficult Workplace Conversations
Effective manager-employee communication is critical for retaining high performers and developing underperformers, yet training managers in these skills remains costly. Text-based chatbots offer a scalable approach but cannot provide realistic rehearsal: managers need to practice speaking aloud to build confidence before high-stakes conversations. In this paper, we propose Conversation Coach, a voice-first AI system that enables managers to rehearse difficult workplace conversations in a realistic spoken format. The system addresses three challenges: achieving low-latency interactions with strong language understanding, enabling adaptive conversations through configurable bot personalities that simulate different employee types, and generating personalized feedback on content and policy compliance. We compare an end-to-end speech-to-speech model with a cascaded approach combining automatic speech recognition, a large language model, and text-to-speech synthesis. The end-to-end approach achieves 3$\times$ lower median (P50) latency with native barge-in capability at an estimated 8$\times$ lower cost, while the cascaded approach offers superior reasoning essential for coaching quality. We deployed the cascaded architecture in production, where 40,000+ managers used it over six months, with adoption patterns indicating selective use for difficult conversations.
Read the full abstract
Human-Computer Interaction
ErgoAssist: Cognition-Aware Posture Feedback in Wearable Ergonomic Systems
Prolonged digital device use has made poor posture and musculoskeletal discomfort pervasive among knowl- edge workers. Existing ergonomic wearables rely solely on posture thresholds, frequently interrupting users during high-focus moments and leading to alert fatigue and abandonment. Yet posture and cognitive load are closely coupled, and most systems remain cognitively unaware. We present ErgoAssist, a head-worn ergonomic assistant that detects poor posture using IMU-based head tracking and estimates task-induced cognitive load using a consumer-grade EEG headband for continuous everyday use. In a controlled lab study, ErgoAssist achieves 81% posture classification and 90.2% task induced cognitive load estimation accuracy under leave-one-subject-out evaluation. In a preliminary real-time deployment, cognition-aware alerting reduces alert frequency by 81%, improves perceived usability by 43%, task performance by 25%, and improves posture correction rate by 38%, delivering fewer but better-timed interventions rather than merely suppressing alerts.
Read the full abstract
Optimization and Control
Models and Algorithms for Reserve Deliverability in Cross-Zonal Balancing Capacity Markets
In power markets where the physics of the transmission grid is closely represented in market clearing models, certain cross-zonal power exchanges can take place only if other exchanges occur concurrently. This leads to challenges in cross-zonal balancing capacity markets, where the activation of reserves in real time remains uncertain. Ensuring reserve deliverability in all activation scenarios is naturally modeled as a stochastic programming (SP) problem, whose size grows exponentially with the number of locations in the network. This formulation scales poorly in real-world applications. We first show that activation scenarios can be expressed in an order-book-agnostic way, reducing the challenge to a network modeling problem and improving computational performance. We then introduce a general inner approximation principle that we use to derive two scalable inner approximations and one column generation algorithm for tackling one of the approximations. The first inner approximation is well known to practitioners, and relates to a basic result for describing boxes in a polytope, while the second model, its associated column generation algorithm and finite-dimensional reformulation are based on semi-infinite linear programming and robust linear optimization. We compare the inner approximations to the exact SP formulation, which we also solve via a Dantzig-Wolfe decomposition for comparison purposes. Numerical results show that these inner approximations are much more scalable than the stochastic programming formulation, while reaping most of the benefits of cross-zonal exchanges. The approaches are of particular interest for future pan-European cross-zonal balancing capacity markets, and can also accommodate co-optimization of energy and balancing capacity products.