A Sharp Dynamic Haussler Lemma (a tight sequential covering bound for learning classes)
WhyProves a sharp global dynamic covering bound for bounded real classes, removing logarithmic overhead and resolving several open questions with complete proofs.
Paper testing how far screening and generative design of crystals can really go
WhyIntegrates 55 databases and completes properties, showing screening is underestimated and inverse design overestimated, backed by controlled comparisons and a released model library.
Quarter-degree probabilistic weather forecasting with a more efficient transport-attention model
WhyProposes two-sided transport attention and self-featurised CRPS, matching GenCast at 9.8% training compute on WeatherBench 2 with broad forecasting impact.
Simulating Driver Interactions with Data-Calibrated Differential Games
WhyCalibrated differential game beats SUMO and inverse RL on held-out vehicles with significance tests and verified approximate equilibria, a strong empirical contribution.
Beyond Perception: A Partition Theory of AI Evaluation and Oversight Limits
WhyProposes a partition-based evaluation theory with a proven one-sided bound and echo-test completeness, substantially advancing theory for AI oversight limits.
Copying as Defense: Stopping Image Edits by Making Models Copy
WhyProposes a novel copy-inducing defense via early attention interventions, substantially cutting edit success across multiple editors and benchmarks.
Online fair division of goods using frequency predictions
WhyGives a complete feasibility boundary for online fair division with frequency predictions, with tight impossibility proofs for five or more agents and rigorous positive results for fewer agents.
Stars are a rough AI assessment based on the title and abstract: each paper is scored for novelty, strength of evidence and breadth of impact, then mapped to 1–5 stars (about 2% get five stars and 10% get four). The site only marks four- and five-star papers as Notable, with a one-line reason; no mark does not mean a paper is weak. The full text is not read, so treat the stars as a reading aid, not a verdict on quality. "Code" means the abstract states that code, data or tools are released.
In briefStrategyproof deep reinforcement learning for allocating online resources
SummaryThis paper addresses two linked problems in online resource allocation: distributing scarce resources efficiently under uncertain future demand, and ensuring agents report task values and requirements truthfully. The authors introduce the first strategyproof deep reinforcement learning mechanism for this setting, combining a monotonic neural network, a generalised critical payment, and training by either RL or imitation of an existing rule. In experiments on a four-resource stream, the neural rule achieves 97.8% of an offline welfare upper bound, and imitation reaches 97.5% in 261 seconds, about 89 times faster than RL.
Fariba Dehghan, Sebastian Stein, Vahid Yazdanpanah et al.
In briefALADDIN: One-Step Face Swapping via Adversarial Latent Diffusion Distillation
SummaryThis paper presents ALADDIN, a one-step face-swapping method that transfers a source identity onto a target image while keeping the target's pose, expression, hair, and lighting. The generator is initialised from a pre-trained multi-step flow-matching model and trained adversarially under identity and target-preservation constraints. The authors report that it is state-of-the-art among one-step methods in realism and target preservation at comparable identity transfer, and matches the best multi-step model with 28 times fewer network evaluations.
Ivan Shchekotov, Timofei Gritsaev, Viacheslav Pirogov et al.
In briefRegression functions as fields over covariates: evolution equations and identification
SummaryThe paper treats the conditional mean of an outcome given covariates at each time as a field over covariate space and studies how it evolves when covariates and outcomes change over time. It shows this field satisfies a second-order evolution equation whose diffusion term is the trace of the regression Hessian weighted by the covariate covariance. The authors also connect the Hessian to quadratic regression and kernel smoother bias, and show that for binary outcomes the equation reduces to a Nagumo-type equation with logistic travelling fronts.
In briefsemanticfa: An R package for analyzing psychological scale structure from item meanings without response data
SummaryThe paper introduces semanticfa, an R package that performs factor analysis on similarities between language-model embeddings of psychometric scale items, without needing respondent data. A worked example applies it to the 50-item IPIP Big-Five markers embedded with Qwen3-Embedding-8B and compares the results with a factor analysis of 874,434 respondents. The authors propose semantic-behavioral coherence, the agreement between semantic and response structures, as a measurable scale property, and argue that semantic factor analysis complements rather than replaces response-based analysis.
In briefTeaching ecological reality to AI: a plug-in training corpus for living-system welfare
SummaryThe paper argues that large language models increasingly make decisions affecting living systems, yet ecological reasoning is rarely a focus of training. The authors built a real-data continued-pretraining corpus covering human, biosphere and AI dependencies, option value under irreversibility, and recovery from harm. Installed on five open models, it reduced harmful choices under the highest pressure to zero or near zero in four of them, with MMLU unchanged. A construct-neutral control corpus showed no such improvement.
In briefFrom Interaction to Capability: Task Advancement as State and Supervision for GUI Agents
SummaryThis paper introduces Task Advancement Memory (TAM), which treats verified task progress as the core mechanism for GUI agents, letting them build on earlier results across applications in long workflows and learn reusable operations. On MobileWorld GUI-only tasks, Advancement State alone raises success from 12.0% to 28.2%, and full TAM achieves a 17.6% relative gain over the strongest baseline while cutting lifecycle tokens per solved task by 38.7%. On SPA-Bench's single-app English Level 3 setting, TAM reaches state-of-the-art task success.
In briefFollowing Gradients to Find Training Data That Erodes Model Safety
SummaryThe paper shows that fine-tuning large language models on instruction data can weaken their safety even when individual examples look harmless. It introduces SimpleLoRIF, a scalable gradient embedding method that identifies training examples whose gradients align with unsafe behavior. Using it, the authors build SafeTulu by removing the 20% of Tulu 3 examples most aligned with unsafe responses, which reduces GCG attack success rate by up to 22% across Qwen, Llama, and Olmo while largely preserving capability.
WhySimplifies low-rank gradient attribution to scale to million-example filtering, cutting GCG attack success up to 22% across several models.
In briefRunning Activity Recognition Inside the Radio: A Spiking Accelerator for mmWave Sensing
SummaryThis paper studies whether spiking neural networks can recognize human activity from mmWave integrated sensing and communication channel estimates while remaining powered when the main processors sleep. The authors implement SNN accelerators with a spike encoder on a ZCU208 RFSoC, so the full chain from micro-Doppler frame to activity label runs in programmable logic. A 132-neuron fully connected SNN reaches 94.7% accuracy on unseen subjects of the DISC dataset with 206 microsecond inference. Compared with conventional networks of the same topology, it uses one-eighth the logic and less dynamic energy per decision.
Anthony S\'anchez, Rafael Ruiz, Jos\'e A. Belloch et al.
In briefHGPRGDel: A Learning and Ranking Method for Growth-Coupled Gene Deletion Design
SummaryThis paper presents a reusable protocol for designing microbial strains in which cell growth is coupled to production of a target compound. It learns from previously computed gene-deletion strategies to rank gene-deletion priorities for each target metabolite, then evaluates candidate strategies with constraint-based metabolic analysis. The learning component, HGPRGDel, explicitly models gene-protein-reaction logic and achieves the highest macro-F1 and AUC among the evaluated configurations on e_coli_core, iMM904, and iML1515. Its rankings recovered an average of 90% of the reference deletion genes from MetNetComp while examining fewer genes than the baseline methods.
In briefQD-PINN: Neural-network method for Bayesian quickest detection in high-dimensional diffusions
SummaryThe paper proposes a physics-informed neural network framework for estimating optimal stopping boundaries in quickest detection problems with high-dimensional diffusion processes. It establishes an equivalence between the free-boundary PDE and an infinite-dimensional stochastic optimization problem to justify using finite-dimensional networks. Uncertainty in the learned boundaries is quantified with deep ensembles and Monte Carlo dropout, showing low variability, and results closely match reference numerical solutions where available.
In briefHow reliable are individual brain-network attributions in ADNI resting-state fMRI cognitive prediction?
SummaryThis paper examines whether per-subject attributions of which functional brain networks drive a resting-state fMRI cognitive prediction model are reliable. Using 875 subjects from ADNI, it predicts MMSE and ADAS-Cog13 from Schaefer-200 functional connectivity. Split-half reliability is near ceiling but falls on scan-rescan retests, and matched pairs exceed a mismatched-subject permutation null. The authors conclude attributions carry a modest person-specific component on top of a strong population-shared pattern.
Sheethal Bhat, Mahfuzur Rahman Chowdhury, Imam Ahmed Shaik et al.
In briefPerceptRL: Making Visual Perception Rewardable in Multimodal LLMs
SummaryThis paper addresses the weakness of fine-grained visual perception in multimodal large language models after post-training. PerceptRL uses a multi-agent pipeline to turn verified errors from dense image captions into QA pairs with rule-based rewards, and adds multi-turn reinforcement learning with visual tools. Coordinate Injection and a grounding-aware process reward correct spatial drift during tool use. Across seven benchmarks, the method reaches 71.6% average accuracy.
In briefReinforcement-learning control for autonomous fixed-wing drone interception
SummaryThe paper presents a fixed-wing interceptor drone controlled by a neural network that maps filtered monocular vision observations directly to actuator commands. The policy was trained in a custom Python environment using reinforcement learning with proximal policy optimization, then deployed on a fixed-wing platform. In simulation it outperformed two classical guidance controllers under most conditions, and it successfully intercepted maneuvering fixed-wing targets in real flight.
In briefWhether later model releases keep earlier pretraining disclosure: a document audit of 41 language model entries
SummaryThis paper audits 41 language model entries from 17 families using 24 pretraining-related indicators from the 2025 Foundation Model Transparency Index. It compares adjacent releases across ten families, tracking which disclosures were gained, lost, or retained. The authors find that scores more often fall than rise, with more indicator losses than gains, and that unchanged or higher totals can hide lost disclosures.
In briefConcordat: a Python library that enforces step-by-step rules on LLM agents at run time
SummaryConcordat is an open-source Python library that models an LLM agent's allowed behaviour as a finite-state protocol enforced during execution. On each turn the model sees only the tools legal in its current state, and illegal calls are rejected before side effects occur. In an evaluation on 50 tau-bench airline tasks with simulated customers pressing for skipped confirmations, the standard agent made unconfirmed changes in several tasks while the Concordat agent made none, with no significant difference in task success.
WhyApplies session types and typestate to enforce protocols on LLM agents at runtime, a novel approach; evaluation is limited to tau-bench airline tasks.
In briefM4Bench: A Benchmark for Procedural Specs in EHR Derivation Agents
SummaryThis paper introduces M4Bench, a benchmark of 28 clinical variable derivation tasks across 15 task families on MIMIC-IV and eICU. It tests whether AI coding agents can reproduce these derivations and how clinician-reviewed agent skills affect their results. Skills raised reward by 0.136 on average, mostly due to specification completeness, while only a few runs passed every strict check. The tasks, skills, evaluation code, and run evidence are released as open source.
WhyIntroduces a new clinical derivation benchmark and decomposes skill value through multiple controlled experiments, offering useful insight for the subfield.
In briefInput Skip Reparameterizes a VAE Without Necessarily Improving Its Generator
SummaryThe paper shows that adding a constant input skip to a Gaussian variational autoencoder can be rewritten so that the distribution of generated samples stays unchanged while reconstruction mean squared error shrinks by a fixed factor. Experiments on a Gaussian case with known density and on seven DART tsunami records find that lower reconstruction error does not reliably improve energy score. The authors conclude that assessing generation requires measures beyond reconstruction error.
In briefInterpretable Meta-Policies for Neural Decision Making: a hierarchical RL framework that puts transparent reasoning in high-level skills
SummaryThis paper presents NEXUS, a hierarchical reinforcement learning framework that combines neural skills with symbolic meta-policies. By moving transparent reasoning to high-level skills rather than atomic actions, it greatly reduces symbolic complexity without sacrificing environment performance. The authors also compare neural, symbolic, and neuro-symbolic meta-policy architectures in terms of environmental return and aligned task performance.
Raban Emunds, Jannis Bl\"uml, Quentin Delfosse et al.
In briefGraph Representation Learning via Adaptive Expectile Aggregation: a learnable expectile level replaces simple averaging to capture more neighborhood information.
SummaryThis paper proposes ExpectileGCN, an adaptive graph aggregation framework that uses a learnable expectile level to capture neighborhood information beyond the mean. The operator moves continuously between minimum- and maximum-type behavior, with mean aggregation as a special case, and is made differentiable through sigmoid smoothing and fixed-point iterations. The authors provide stability and transductive generalization guarantees and analyze classification risk under heterophilic neighborhood models. Experiments on homophilic and heterophilic benchmark graphs show strong predictive performance and improved robustness under feature corruption and structural perturbations.
In briefTest-Time Communication in Trading Agent Swarms: How Information Sharing Among Independent Trading Agents Affects Holdings Convergence
SummaryThe paper studies 200 agents from four model families, grouped in tens, trading under five communication regimes over seven US trading days in October 2026. In two families, communicating groups held far more of the same stocks than isolated groups, and group returns largely reflected shared holdings. Because isolated groups also converged, the authors argue herding should be measured as excess convergence against a no-peer baseline.
Avid Fayaz, Harry Zhang, H{\aa}vard Alstadheim et al.
In briefETAMem: A Memory Framework for Long-Term Chatbot Conversations
SummaryETAMem is a multi-view memory framework for long-term conversational agents built on large language models. It breaks conversation information into atomic memory units, identifies affective content, and organizes these units into three complementary views: Event, Affect, and Time. Retrieval allocates budgets across views based on each query. On the LoCoMo and LongMemEval-S benchmarks, it reports relative accuracy gains of 12.46% and 7.96% over state-of-the-art baselines.
In briefCRISP: Co-designing Pruning and Scheduling for Efficient MoE Inference on Edge Servers
SummaryThe paper addresses efficient inference for Mixture-of-Experts models on memory-constrained edge devices, where moving expert weights dominates latency. It proposes CRISP, which reorders neurons by importance so one stored expert supports multiple execution widths, and assigns differentiated execution to hot, warm, and cold expert combinations based on frequency and expert role. Evaluated on Mixtral-8x7B and DeepSeek-MoE-16B, CRISP improves throughput by up to 2.1 times at the same memory budget, with accuracy loss limited to 3.6% relative to full-width inference.
In briefRethinking Visual Imagination for Spatial Reasoning: Representing Unseen States Beyond RGB
SummaryVisual world models often generate RGB views of unobserved states to support spatial reasoning, though RGB is mainly the default because it suits visual generation. The authors introduce ImagiRep, a simulator-grounded benchmark of 922 questions across seven types and two settings, offering RGB and six alternative representations such as depth, panoramas, and top-down geometry. They find that accurate imagined RGB does not ensure accurate reasoning, and different questions benefit from different representations. They also propose GEARS, a lightweight test-time method for adaptive representation selection, which outperforms existing spatial imagination methods.
In briefFeatureArc: Testing Coding Agents on Features That Grow Over Time
SummaryThis paper introduces FeatureArc, a benchmark that evaluates coding agents on software features that evolve through successive stages, with each extension building on the agent's own earlier code. Tasks are built from repository histories and checked with executable tests across four programming languages. Evaluating 10 agents on 73 tasks, pooled success falls from 62.7% at the first stage to 28.5% on the whole feature. Inherited defects from earlier stages are a common cause of failure.
WhyProposes a benchmark and pipeline for evolving features; experiments on 10 agents and 73 tasks reveal a large first-stage versus whole-feature success gap.
In briefConfidence Scoring for Segmentation in Any CNN
SummaryConvolutional neural networks are widely used for medical image segmentation but lack a built-in confidence measure. This study explores voxel-based uncertainty estimation methods that can be added to existing segmentation pipelines without training from scratch. It also proposes a new method for deriving a single confidence score from uncertainty maps, and evaluations on different regions of interest and modalities suggest this score can predict segmentation performance on unseen data.
In briefHow instructions, post-training, and internal activations affect whether multimodal models refuse questions a picture cannot answer
SummaryThis paper studies whether multimodal language models refuse to answer questions that an image cannot support. Across three models on the MoHoBench benchmark, the authors compare inference-time instructions, supervised and preference-based post-training, and interventions on a refusal-related activation direction. Refusal is sensitive to how instructions are worded and placed, post-training raises refusal more strongly, and activation interventions produce smaller changes that transfer across prompt conditions.
In briefAdaptive rollout allocation to improve GRPO training for LLMs
SummaryThe paper notes that GRPO gives every training question the same rollout budget, ignoring differences in difficulty. It proposes AdaRoll-GRPO, which models rollout allocation as a Bayes-adaptive Markov decision process to learn an adaptive allocation policy. The policy is computed offline and applied during training, under the same total rollout budget as standard GRPO. Experiments on multiple reasoning datasets and LLM architectures show improved GRPO training performance under a fixed budget.
Faeze Moradi Kalarde, Armin Toroghi, Nam Ho Koh et al.
In briefShared Latent Number Geometry Across Text and Image Arithmetic
SummaryThis paper asks whether a multimodal model uses one internal concept of number for arithmetic given as text or as an image. Using causal interventions on Gemma4-12B-it's residual stream, the authors learn subspaces where swapping coordinates changes the model's answer. The subspaces for different conditions use different directions but can be aligned by one rotation and rescaling. The study also finds hidden directions that carry the number without affecting the current output digit and that steer later digits.
WhyUses causal interventions rather than similarity metrics to show text and image arithmetic share one numerical geometry, with hidden directions carrying the number.
Oriol Juan Sabater, Alex Ferrando de las Morenas, Pau Rodr\'iguez et al.
In briefDisplacement Moments and Path-Terminal Separation in Brownian Interpolant Sampling
SummaryThis paper studies sampling error in Brownian stochastic interpolants, which connect paired endpoint samples, and shows that this error can depend on displacement moments even when neither endpoint has a first moment. For arbitrary Borel couplings, the authors prove early-stopped KL guarantees from target to output for standard Euler sampling. The work also compares bridge-path, comparison-path, and terminal KL behavior, and identifies the source information lost through Markov projection.
In briefCTBakeGS: Converting photorealistic CT renderings into 3D Gaussian splats without training
SummaryThis paper presents CTBakeGS, a method that converts photorealistic CT volume renderings directly into 3D Gaussian splatting (3DGS) models without any training. Splats are placed on an adaptive octree of millimetre-sized cells, with colour and opacity computed on the graphics processor. Tested on six public CT scans, two-million-splat models reached PSNR values of roughly 30 to 33 dB, and conversion took seconds to about a minute on ordinary computers. The converted models rendered at 68 frames per second on a laptop.
Haruki Fukuda, Constantin Lux, Wolf Schweitzer et al.
In briefClearing the FoG: training knowledge graph agents to retrieve evidence for their answers
SummaryThe paper argues that existing knowledge graph question-answering agents are trained and rewarded only for correct answers, so they often answer correctly without retrieving the supporting triples. It proposes FoG, which combines evidence-grounded question synthesis, a KG interaction agent, and an evidence-coverage process reward that credits each supporting triple to the turn that first retrieves it. On three KGQA benchmarks, 3B–4B open-source agents trained with FoG outperform substantially larger baselines in answer accuracy and evidence coverage, and transfer zero-shot to other knowledge graphs.
In briefAn end-to-end framework for validating generative models with uncertainty
SummaryThe paper proposes SIGMA, an end-to-end framework that combines density estimation with principled uncertainty quantification. Using Gaussian mixture models, the authors profile fitting uncertainties as nuisance parameters in a composite goodness-of-fit test, rather than treating the model as exact. Experiments on a two-dimensional toy model and a collider physics QCD background benchmark show the model can be trusted on test sets up to an order of magnitude larger than standard tests allow.
Giada Badaracco, Sean Benevedes, Christina Reissel et al.
In briefGame-theoretic agent with Bayesian opponent modeling and moral rhetoric for the GLEE competition
SummaryThis paper describes an agent submitted to the GLEE Competition, which involves bargaining, negotiation, and persuasion games against human and AI opponents. The agent combines a classical game-theoretic solver for numeric moves, an online Bayesian layer that estimates opponent types, and a language model that writes messages. It was evaluated over 13,200 Monte Carlo games against 11 synthetic opponent archetypes and through analysis of 9,337 live-game messages. In the persuasion setting, moral rhetoric framing was strongly correlated with whether the buyer purchased, even after controlling for quality, round, and configuration.
In briefLearning Robot Policies Whose Execution Quality Can Be Controlled via Iterative Hindsight Relabeling
SummaryThe paper introduces Quality-Controllable Policy (QCP), a robot policy whose execution speed, motion smoothness, and their combinations can be controlled at inference time. To expand the controllable quality range beyond the initial demonstrations, it adds an iterative loop that collects rollouts under target quality labels, relabels trajectories by their measured quality, and fine-tunes the policy. Experiments on Robomimic and RoboEval show improved execution quality and success rates across iterations, and real-world tests show gains in conveyor throughput and wine-glass transport success.
In briefJEPA-inspired latent field prediction for reconstructing MIMO-OFDM channels from sparse pilots
SummaryThe paper addresses the high pilot overhead needed to acquire full channel state information in MIMO-OFDM wireless systems. It proposes a JEPA-inspired world model of the wireless propagation field that maps channel observations into latent time-frequency tokens and predicts masked or future latent states from historical multi-band context. The predicted prior is fused with sparse noisy current pilots to reconstruct the full channel. Ray-tracing experiments show improved normalized mean square error over pilot-only and direct reconstruction baselines.
In briefLinMix: Data Mixing Framed as Linear Domain Attribution
SummaryThis paper studies how the mixture of domains in a language model's pretraining data affects performance. It proposes LinMix, which samples data mixes at the vertices of a capped simplex so each domain is at its cap or absent, yielding an interpretable linear model whose coefficients are domain attributions. In pretraining experiments with 1B and 4B parameter models, LinMix outperforms Dirichlet-based mixing on reasoning, code, and math benchmarks, and it supports linear fitting across 288 fine-grained domains.
WhyRecasts data mixing as linear domain attribution, replacing Dirichlet with vertex sampling and proving optimality; evaluated on two models and several benchmarks, giving fairly solid evidence.
Hubert Baniecki, Przemyslaw Biecek, Martin Pawelczyk
In briefHow many and how much negative samples? RADAR, a contrastive learning method for knowledge graph embedding
SummaryThis paper examines how many negative samples to use for each positive sample and how much each should contribute when training knowledge graph embeddings with contrastive learning. The authors propose RADAR, which combines risk-aware negative budget allocation with distribution-calibrated negative weighting. Experiments on 3 real-world knowledge graphs with 4 embedding models show RADAR achieves the best average performance across all combinations and reduces sampled false negatives by over 90%.
In briefPara-Gaussian approximations for random deep neural networks, a more accurate model of their behaviour beyond the Gaussian limit.
SummaryRandomly initialized neural networks are known to behave like Gaussians as width goes to infinity, but this approximation is poor at finite width and worsens with depth. The paper introduces a slightly non-Gaussian, more accurate approximation called the para-Gaussian distribution, built from Hermite polynomials. It gives a practical recursive method for approximating deeper layers. The theory is applied to linear networks, where results are exact, and to ReLU networks, where they are approximate.
In briefCalibration-free channel selection for Mixture-of-Experts decoding
SummaryThis paper studies how to choose which intermediate channels of Mixture-of-Experts models to read during decoding. It finds that calibrated thresholds neither fix the read budget nor reliably improve accuracy, while a top-k selection per token needs no calibration data and matches or beats threshold versions in most comparisons. The authors also propose a reader built only from the model's weights, which retains 98.1% of full-model accuracy while reading 17.8% of routed-expert weights. Triton kernels are reported to speed up MoE blocks compared with optimized vLLM.
In briefPDRVLA: A Vision-Language-Action Model with Physical Dynamics Representations
SummaryThis paper presents PDRVLA, a model that generates robot actions from visual observations and language instructions. A lightweight Physical Dynamics Module extracts physical tokens from vision-language features to condition actions without online video generation, reducing control-loop latency. The model reports success rates of 99.3% on LIBERO, 89.9% on LIBERO-Plus, 93.6% on RoboTwin 2.0, and 81.5% on SimplerEnv across four simulation benchmarks.
In briefSame Verdict, Different Readers: Tailoring Fact-Check Explanations to Individual Preferences
SummaryThis paper studies how to adapt an existing fact-checking explanation to individual readers while keeping its factual content intact. Instead of using reader-written references, the authors learn the rewriting from relative preferences of LLM-simulated readers. They propose PERPL, a framework with simulated preference construction and progressive preference learning, which discovers preferred explanations beyond predefined styles. The experiments report that PERPL outperforms much larger general-purpose LLMs on personalized preference alignment while maintaining high factual fidelity.
In briefAgent-Conditioned Routing for Diverse Reasoning in a Frozen Model
SummaryThis paper asks whether one frozen sparse mixture-of-experts model can produce diverse solution methods for a single problem without changing its weights. The proposed Agent-Conditioned Routing assigns each parallel trajectory persistent expert-eligibility masks while keeping the learned token router. On a 126-question math and physics benchmark, it yields more problems with at least two distinct primary methods than temperature sampling, with somewhat lower correctness.
Ananth Kalyanasundaram, Ankush Kadu, Athira Gopal et al.
In briefEuclidean-crowding-driven adaptive Euclidean-hyperbolic graph learning: a paper analyzing how Euclidean and hyperbolic spaces differ on hierarchical graph data and proposing a method that fuses both geometries.
SummaryThe paper geometrically quantifies Euclidean crowding, showing that fixed-dimensional Euclidean embeddings of tree-like graphs suffer layer-wise minimum separation shrinkage that weakens margin-based generalization bounds. It also provides a generalization bound explaining the advantage of hyperbolic geometry on hierarchical data. The authors propose EH-Fusion, which lifts Euclidean GNN representations into hyperbolic space via exponential mapping and adaptively fuses both geometries based on node-level neighborhood expansion patterns. Experiments on benchmark datasets are reported to validate the theory and the method.
In briefGeometry of manifold learning: singular foliations from deep learning models
SummaryThis theoretical paper studies the geometric structure a trained neural network defines on its data space, using the Data Information Matrix, a variant of the Fisher Information Matrix. The authors focus on the non-smooth case of ReLU activations, where traditional tools often fail. They prove that the distribution spanned by the matrix's columns is non-smooth in the distributional sense, and introduce a smoothed version differing only on a measure-zero set. They show this smoothed distribution induces a singular foliation on the training data space.
Gabriele Novelli, Eliot Tron, Emanuele Latini et al.
In briefWhy LLMs Fail: Tracing Reasoning Failures Inside the Model
SummaryThis paper introduces CauTrace, a white-box diagnostic framework that examines how large language models fail during reasoning, using seven functional states, representation propagation, temporal-state coordination, and reasoning trajectory topology. It is evaluated on clean reasoning failures and paired adversarial perturbations across Qwen and LLaMA-family models. The failed traces show weaker Attention-MLP agreement on support-state updates and less verification relative to commitment, with exact signatures varying by architecture.
In briefTRIDENT-Seq: Label-Efficient Payment Risk Modeling via Profile-Matched Prefix-Future Contrast
SummaryThis paper presents TRIDENT-Seq, a transaction foundation model pretrained on unlabeled payment histories using Prefix-Future Contrast (PFC), which matches each prefix to summaries of subsequent behavior among candidates with similar activity profiles. Across three production payment businesses, the pretrained encoder improves average precision and recall at a 0.5% intervention budget over scratch training with the same architecture. In a deployment evaluation at Ant International, recorded amount-based chargeback and overall loss rates are 37.7% and 37.9% lower than those of the previous XGBoost-based system.
WhyProposes profile-matched prefix-future contrast beating random candidates; validated across three production businesses with label efficiency gains and deployment loss reductions.
In briefEmbeddingGemma 2: A Unified, Efficient Model for Multimodal Embeddings
SummaryThis paper introduces EmbeddingGemma 2, an open model that maps text, images, audio, video, and their combinations into one shared embedding space. It uses simplified per-layer embeddings and a multi-stage training curriculum that gradually adds modalities and objectives. The authors report that it outperforms specialist models and models two to twelve times larger across embedding benchmarks, and it supports flexible deployment options for resource-constrained settings.
Henrique Schechter Vera, Sahil Dua, Jovan Andonov et al.
In briefOnline interval selection with binary predictions: optimal consistency and robustness tradeoff
SummaryThis paper studies online interval selection where each request comes with a binary prediction of whether it belongs to a final optimal solution. Requests arrive in arbitrary order, and accepted requests may be revoked, but rejected or revoked requests cannot be recovered. The authors show that the best consistency under finite strict robustness is (3+√5)/2, and give a deterministic algorithm achieving this consistency with exact robustness φ^5. They also present a continuous family of algorithms with a tunable replacement cost, and show that randomization allows expected consistency approaching one with finite expected robustness.
In briefTimeLineXR: Interactively Exploring How Vision-Language Models Encode Time
SummaryThis paper presents TimeLineXR, an interactive demo that turns the time axis inside vision-language models into a 3D scene. Users can upload images on a web page or in a WebXR environment and see each placed on a color-coded time axis. The system lets users compare where objects land and read their estimated first year of appearance.
Nidham Tekaya, Markus Passecker, Matthias Zeppelzauer
