In briefSwarm Telepathy: Sharing Model Memory Across Different LLM Families Without Re-Reading Text
SummaryThe paper introduces SwaT (Swarm Telepathy), the first method to transfer KV-cache between LLMs from different model families, so the receiver does not need to prefill text. Combined with KV-cache compaction, it outperforms text-to-text communication, averaging 40% higher accuracy and about 4 times faster wall-clock time. Tested on 17 benchmarks across 12 LLMs from 6 families, it can also pass a vision-language model's visual capabilities to a text-only LLM. Both models stay frozen, and training needs under 50M tokens and less than 1 GPU-day.
WhyFirst cross-family KV-cache transfer, validated on 17 benchmarks across 12 LLMs with 40% mean accuracy gain and 4x speedup.
Abdulrahman Alabdulkareem, Irene Tenison, Yang Zhang et al.
In briefTurnCost: Turns, Not Observations — Cache-Aware Loop Engineering for Coding Agents
SummaryThis paper analyzes 600 public coding-agent trajectories and finds that tool observations are usually short, while about 98.6% of input tokens are served from a prefix cache, so flat input pricing overstates cost roughly sevenfold. It reports that turn count explains far more of the variance in trajectory cost than average observation size, suggesting unnecessary turns may be a larger optimization opportunity than compressing observations. The authors do not show that turns can be removed without hurting task success. They also release TurnCost, a tool for re-scoring trajectories under different caching, pricing, compression, and routing policies.
In briefVLCQ: Post-training quantization for deep neural networks using variable length coding
SummaryThis paper proposes VLCQ, a post-training quantization method that replaces fixed-point encoding with variable-length encoding to reduce bit-widths. It requires no training data or fine-tuning. Experiments on image classification, segmentation, and object detection report improved compression and accuracy over prior methods. The authors report up to 12.25x compression versus FP32 with less than 4% accuracy degradation.
Reem Abdel-Salam, Ahmed H. Abdel-Gawad, and Amr G. Wassal
In briefNeuro-symbolic Reasoning for Detecting Persuasion in ChangeMyView Debates
SummaryThis paper proposes NARS, a neuro-symbolic system that detects persuasiveness directly from online discussions. Its neural component uses retrieval-augmented iterative LLM reasoning to extract bipolar argument structures, while its symbolic component applies gradual argumentation semantics to propagate influence across them. Built on the Webis-CMV-20 corpus via the new CMV2BAS dataset, NARS achieves F1=0.57 and AUC=0.56, outperforming zero-shot Qwen3.6:27B and GPT-oss:20B.
Siddharth Bhargava, Sara Tonelli, Patricia Mart\'in-Rodilla
In briefA self-evolving visual agent that learns tool use from unlabeled images using a margin reward
SummaryThis paper proposes Marvo, which lets visual agents learn to invoke tools from unlabeled images without human-annotated data. It uses a challenger agent that creates question-answer pairs and a solver agent trained with reinforcement learning. The challenger is rewarded by the solver's performance gap with and without tools, forming an adaptive curriculum. Across several benchmarks, Marvo improves average accuracy by up to 7.1%.
In briefCan language models learn while reading? Testing gradient updates across architectures
SummaryThe paper tests dynamic evaluation, where a language model takes gradient steps on the text it reads. On softmax-attention models, this lowers next-token loss by about 5-8% over a document and beats the same frozen model given the full context. Hybrid linear-attention models gain under 1%, while masked-diffusion models show larger gains. On long-context benchmarks, however, weight updates recover none of the accuracy lost to truncation, suggesting the adaptation captures style and vocabulary rather than retrieval.
In briefTransferring knowledge between simple grid puzzles to test how models reason under unfamiliar rules
SummaryThis paper studies whether a model can adapt to new rules in small discrete grid worlds, which are Sudoku-like puzzles built from a shared base and extra rules. A single depth-recurrent model was trained on many rule combinations and tested at four novelty levels, with rules given either as a symbolic description or as two solved examples. For recombining rules seen in training, the description scored 99% versus 95% for examples. For rule components that never co-occurred in training, the description fell below the no-information baseline while examples stayed above it.
In briefRegion-aware Squeeze-and-Excitation networks for image segmentation
SummaryThe paper modifies Squeeze-and-Excitation modules by replacing global average pooling with spatially guided pooling, so channel recalibration is region-aware. It proposes four modules (BASE, SPASE, MAPSE, PLEASE) that define regions via a regular grid, SLIC superpixels, semantic maps from a frozen network, or learned soft assignments. The authors report that these modules outperform SE and related attention variants on most benchmarks across segmentation and classification tasks.
In briefTheoretical limits of learning unconstrained machine-learned interatomic potentials
SummaryThis paper examines whether matching energies and forces alone is enough for machine-learned interatomic potentials to sample molecular dynamics accurately. For unconstrained hypothesis classes, it shows that the link between prediction error and sampling accuracy is exponentially ill-conditioned in dimension. It identifies two failure modes, kinetic trapping and equilibrium mismatch, and establishes global energy-control conditions that guarantee equilibrium accuracy.
WhyProves exponential ill-conditioning between MLIP prediction error and sampling accuracy, identifying kinetic trapping and equilibrium mismatch as two distinct failure mechanisms.
In briefBrainAid: Lightweight CNN and Explainable AI for Efficient Brain Tumor Detection
SummaryThis paper presents BrainAid, a system that combines CNNs and Vision Transformers to classify MRI scans into four classes: glioma, meningioma, pituitary tumor, and no tumor. The authors report 98% precision, F1 score, and accuracy on a public dataset, and add Explainable AI to improve interpretability. The system is deployed on a Raspberry Pi, showing that real-time, edge-based classification is feasible in low-resource settings.
Sai Sriram Gonthina, Ross Gore, Bikash Chandra Singh et al.
In briefA multimodal AI model that rebuilds plasma conditions from incomplete measurements
SummaryThis paper presents a proof-of-concept multimodal foundation model for reconstructing plasma states during magnetic reconnection from sparse, incomplete diagnostics. Trained on 250 synthetic VPIC simulations, a 3D U-Net reconstructs three magnetic-field components and electron density and completes missing data across modalities. Within this simulation regime, magnetic and density information improve each other's reconstruction. Generalization to other plasma regimes or experimental measurements remains untested.
In briefGlobal disaster risk assessment using World Risk Index data with geospatial and visual analytics
SummaryThis study analyzes World Risk Index (WRI) data from 2000 to 2025, a metric reported by the Humanitarian Data Exchange that measures national susceptibility to natural disasters. The WRI combines susceptibility, vulnerability to natural hazards, and coping capacity. Using Matplotlib, Seaborn, and Geopandas for visualization, the authors found high WRI values for the Philippines, Mexico, Indonesia, and India, and argue these countries should prepare for disasters.
In briefOffering AI inference as a built-in 5G network service
SummaryThis paper proposes Inference on Demand (IoD), which exposes split AI inference through a CAMARA-like API and carries split-aware tensor context in-band over the 5G user plane. The authors evaluate IoD on a programmable UPF-based testbed using a partitionable visual AI workload. They report that adaptive split selection can run as a network-facing service while avoiding the rigid cost of static configurations.
Rodrigo Moreira, Larissa Ferreira Rodrigues Moreira, Joberto S. B. Martins et al.
In briefComparing Four Deep Learning Models for Segmenting Dental X-ray Images
SummaryThis paper compares four semantic segmentation models, PSPNet, DeepLabv3, UNet, and UNet++, under controlled settings on 1269 dental radiograph images. The models were evaluated using Dice coefficient, IoU, accuracy, precision, and training and validation losses. UNet++ achieved the best results, with an IoU of 0.6719 and a Dice value of 0.7869, at an image size of 300x300 and a batch size of 16. The study also examines how image resolution and batch size affect model performance.
In briefThe Optimal Decoder is Tilted Temperature: a decoding rule for LLM sampling that balances diversity and plausibility
SummaryThe paper frames LLM decoding as maximizing pass@k under an entropy floor and a cross-entropy ceiling, deriving a closed-form "tilted temperature" sampling rule. Under a stated scaling of the diversity budget with sample size k, the optimal temperature is shown to increase with k. It introduces J-EOS for online learning from pass/fail feedback and Calibrated Temperature Decoding (CTD) for offline use, and reports gains over standard and adaptive baselines on HumanEval and MBPP.
Nikolaos Papagiannis, Ananth Grama, Wojciech Szpankowski
In briefChoosing or Rewriting: Oracle-Guided Fine-Tuning for Generative ASR Post-Editing from N-Best Lists
SummaryThe paper studies how large language models can correct speech recognition output by comparing N-best hypotheses and rewriting the transcript when needed. The authors fine-tune models on three curated pools: corrections that teach selection, top-1-correct cases that discourage unnecessary edits, and hard negatives that reduce hallucination. Gains are clearest around 3B parameters on a proprietary dataset, while results on CommonVoice are less conclusive.
In briefTrace Evolution Bench: Testing Whether Agents Improve from Their Own Environment Traces
SummaryThis paper introduces Trace Evolution Bench (TEB), a protocol that separates genuine agent self-evolution from simple frozen rescoring, along with three frozen data packs. On the coding pack, DeepSeek v4 Pro under no_evolve never changes its policy yet varies by 7.5, 5.0, and 12.5 points across three seeds. Text-update methods show mixed results across models, and no method stays above its own seed variation on every seed of all three models.
In briefUsing Convolutional Neural Networks to Detect Branch Dieback in Pequi Trees
SummaryThis study compares five convolutional neural network architectures (AlexNet, Inception, LeNet, VGG16, and MobileNet) for diagnosing branch dieback in pequi trees in the Brazilian Cerrado. The models were trained on 6,272 multi-scale images captured by UAV and handheld sensors under uncontrolled lighting. Results show that greater architectural complexity did not ensure reliability, while MobileNet performed best with a Precision of 0.71 and ROC-AUC of 0.63. The authors suggest lightweight CNNs may be viable for automated plant health surveillance in resource-limited field settings.
Ronilson Rodrigues Pinho, Elua Ramos Coutinho, Nilza de Lima Pereira Sales et al.
In briefBelief Embedding Tree Search: a Monte Carlo planner using learned belief representations for partially observable decisions
SummaryThis paper addresses planning in partially observable Markov decision processes (POMDPs), where beliefs are posterior distributions over hidden states and exact planning is generally intractable. The authors propose Belief Embedding Tree Search (BETS), which represents each node's belief with a generative distribution embedding updated through the POMDP dynamics and trains this belief model alongside policy and value networks. On six configurations of RockSample, LightDark, and PocMan, BETS is competitive or leading against BetaZero and three planners without learned components.
Christopher Solinas, Radovan Halu\v{s}ka, Martin Schmid et al.
In briefPAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media
SummaryThis paper proposes PAT, a framework for automatically detecting depression from social media posts. It first trains a post-level encoder using fine-grained contrastive learning with LLM-based data augmentation grounded in classic personality theories, then uses the trained scorer to filter high-quality posts and transfers the encoder to user-level tasks. The authors report that PAT significantly outperforms existing methods and provides interpretability by tracing mental-state trajectories and highlighting key symptoms.
In briefTheory of online temperature sampling for language model decoding
SummaryThe paper addresses repetitive, degenerate text from LLM decoding by maximizing sampling entropy under a single cross-entropy budget. It studies the online setting, where conditional distributions arrive one at a time and the temperature must be chosen causally. The proposed Entropic Online Sampling (EOS) method loses O(R log n) entropy relative to the offline optimum and is shown to be optimal up to a log n factor. Experiments indicate EOS tracks the offline temperature and achieves higher entropy than fixed-temperature sampling.
WhyExtends offline temperature sampling to an online setting with near-optimal loss and budget bounds, a substantive theoretical advance, though experiments are limited in scale.
Jin Sima, Nikolaos Papagiannis, Ananth Grama et al.
In briefSparse routing saves compute, not samples: error bounds for mixture-of-experts models
SummaryThis paper studies the estimation error of least-squares Mixture-of-Experts (MoE) models whose experts and gating networks are Lipschitz in their parameters. It derives an oracle inequality with a fast rate for dense MoEs and extends it to sparse Top-k and Top-p routing, where an extra cost governed by the Natarajan dimension of the selector appears. For linear experts, the authors prove a matching minimax lower bound, showing the rates are optimal up to logarithmic factors.
In briefWhat actually improves universal multimodal embeddings: a controlled comparison of contrastive training methods
SummaryThis paper systematically compares several contrastive learning training strategies for multimodal embeddings under a unified benchmark, MMEB. Most sophisticated strategies yield only marginal gains over a standard baseline, and some reduce overall performance. The authors use an MLLM-as-a-Judge diagnostic to analyze retrieval errors and report that combining hard-negative mining, false-negative mitigation, and hard-batch construction gives the best results, averaging a 6.0% improvement with Qwen2-VL-2B.
In briefWhat Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization
SummaryThis paper studies why generative search engines such as ChatGPT, Gemini and Perplexity cite particular web pages, using two 90-day collections totaling 158,099 answers and 468,037 citations to commercial questions. Content that matches the questions people ask is cited more often within the same website, but page properties explain little of the difference between pages. The authors find that the engine's own retrieval step and the model's prior knowledge of brands explain much more of the variation, and they release their data and code.
In briefInCLAD: A Continual Learning Benchmark for Industrial Visual Anomaly Detection
SummaryThis paper introduces InCLAD, a continual learning benchmark for industrial visual anomaly detection. Each industrial component is treated as a task, and models learn semi-supervised detectors sequentially while needing to keep performance on earlier components. The benchmark defines random, easy-to-hard, and hard-to-easy task orders, plus scenarios spanning multiple datasets. Performance is evaluated at both image-level detection and pixel-level localization.
In briefAppearance Compensation in Volume Rendering: Matching Renders Does Not Guarantee Correct Density
SummaryThe paper asks whether a volume can reproduce every external RGB view while still having an incorrect density in emission-absorption rendering. It combines theoretical results on optimal density gain and identifiability with a constructed physical SH2 example. In that example, central density doubles and relative density L2 error exceeds 14.52% while every external RGB ray is unchanged. The authors conclude that rendering accuracy alone cannot certify density, since recovery also depends on appearance capacity, spatial resolution and measurement design.
WhyProves density is not uniquely recoverable from RGB rendering under appearance compensation, with an explicit exact physical SH2 ambiguity construction and controlled experiments.
In briefIs a 9B decision model needed for symptom-to-disease classification? Clef-Flash fine-tuning versus TF-IDF baselines
SummaryThe paper fine-tunes Clef-Flash, a 9-billion parameter decision model, with QLoRA to classify 24 diseases from symptom descriptions in a single forward pass. The fine-tuned model reaches 95.42% accuracy, while TF-IDF linear SVM and logistic regression baselines trained on the same split reach 98.75% with fast CPU inference, so no accuracy advantage is shown on this benchmark. The authors conclude that validation on longer, real clinical text and clinician review are needed.
In briefUltraBERT: Pushing the Limits of BERT Pre-Training
SummaryThis paper revisits the training choices for bidirectional encoders and finds that performance on downstream tasks saturates beyond a few trillion tokens. It reports that lower masking ratios help classification while higher ones help retrieval, and that distillation with pruning has opposite effects on the two. Based on these findings, the authors propose UltraBERT, which uses separate LM heads for low and high masking ratios and applies distillation only to the low ratio. The model reaches 91.0 on GLUE and 47.7 on BEIR, which the authors say makes it the first encoder to be state of the art on both classification and retrieval.
In briefPATTERN Can Be Solved from Memory: Generator-Aware Baselines for Synthetic Node Classification
SummaryThis paper examines PATTERN, a synthetic node-classification benchmark used to rank graph networks, whose generator reuses the same 100 patterns across all splits. A training-free matcher that reads patterns from training labels reaches 99.2 class-weighted accuracy, above the best published model at 89.8. The authors also find that CKGConv and GRIT score below a pattern-blind predictor on fresh patterns, and recommend reporting scores on new templates and publishing generator-based baselines.
In briefAlgorithmic Fairness Over Time: Advances and Prospects. A paper on measuring and ensuring fairness of algorithmic decisions over time.
SummaryThe paper explains that when feedback loops exist, an algorithm's decisions at one time step can influence later observed data, potentially causing selection bias or distribution shifts. Ignoring these effects can produce unfairness that may worsen as the model stays deployed. The authors present a taxonomy distinguishing static and temporal fairness notions from recent work, then identify common implicit temporal assumptions and open questions for future research.
In briefCowCare: A Unified Framework for Tracking Cattle Across Views and Analyzing Heat Stress
SummaryThis paper presents CowCare, a framework that tracks individual dairy cows across multiple farm cameras and recognizes heat stress behaviors in their videos. Its tracker, CowTrack, keeps cow identities consistent across camera views using a view-decoupled network, while TSDF-Net, a video model, detects heat stress using temporal stabilization and diffusion-based semantic priors. On two real farm benchmarks, the authors report 83.48% HOTA and 87.72% IDF1 for tracking on MVCow, and 87.72% Top-1 accuracy on HSCow.
In briefBudget-constrained sequential decisions with provable sublinear regret
SummaryThis paper studies automated decision makers, such as AI agents controlling robots or network devices, that must keep a cost measure within a per-episode budget while staying competitive in reward. The authors propose the online conformal episodic agent (OCEA), which picks one candidate action per episode with a regret-controlled algorithm and switches to a zero-cost action when the remaining budget cannot cover another round's maximum cost. They prove that OCEA meets the per-episode cost constraint for every state sequence and randomness realization, and achieves sublinear expected policy regret against the best hindsight action sequence from a comparable constrained class. Applications include risk-averse control and online compression with distortion guarantees.
Jayadev Naram, Giuseppe Durisi, Matteo Zecchin et al.
In briefComputational complexity gap between POMDPs and multi-environment POMDPs
SummaryMulti-environment POMDPs (ME-POMDPs) model uncertainty by encoding a finite set of possible environments. Prior work showed that the finite-horizon robust reward maximization problem for ME-POMDPs is PSPACE-complete. This paper identifies which subclasses keep this complexity, proving NP-completeness for multi-environment blind MDPs and deterministic POMDPs, and PSPACE-completeness for several subclasses lacking a closure-under-mixture condition.
WhySystematically maps complexity boundaries of ME-POMDP subclasses, giving new NP, PSPACE, and existential theory of reals results with rigorous proofs.
Eline M. Bovy, Caleb Probine, Marnix Suilen et al.
In briefCan We Trust Preference Models? A Stackelberg Game for AI Alignment
SummaryLanguage models are often aligned to human preferences using imperfect evaluators whose errors can be exploited during optimization. The paper trains policies to find comparisons where an evaluator may reverse the true preference, using disagreement among a preference-model ensemble as a proxy for prediction error. This search is formulated as a zero-sum Stackelberg game, with equilibrium results and an algorithm, SMIDA, that converges linearly in the tabular setting. Experiments with a synthetic oracle show the trained policies find more held-out comparisons exceeding the disagreement threshold and reversing the oracle's preference than reference policies.
In briefDisCo Algebra: Why Splitting and Recombining Probabilities Are Not Inverses
SummaryThis paper introduces DisCo algebra, a compositional framework for joint probability built from a disintegration operation and a non-associative combination operation. Disintegration summarizes each block of variables with joint CDFs, which compresses information, while combination reconstructs distributions using regular conditional measures. The authors prove that combining after disintegrating is the identity on admissible inputs, but disintegrating after combining is not. The information lost in that case is quantified as the probability curvature, the conditional KL divergence between the true conditional distribution and its conditionally independent product.
In briefPruning Vision Transformer layers without retraining to speed up inference
SummaryThe paper measures cosine similarity between layer inputs and outputs to find redundant layers in vision transformers, finding that deep models such as DINOv2 and SwinV2 show significant redundancy. On ImageNet-1K, removing 20% of DINOv2 blocks without retraining gives a 1.24x throughput gain and 20% lower energy per image, with under 1% relative accuracy loss. The authors also propose a pipeline that automatically selects the best pruned configuration for a given dataset.
In briefDiffGA: Guiding Frozen Diffusion Models Toward Materials Goals via Evolutionary Search
SummaryThe paper presents DiffGA, a method that adds evolutionary search to the sampling process of a pretrained diffusion model, so it can target any black-box objective without retraining. To reduce costly property evaluations, it uses a Gaussian process surrogate built on machine-learned force field embeddings. On three crystal-design tasks with a frozen MatterGen backbone, DiffGA reports higher rates of stable, unique, and novel hits than several baselines. The authors also use density functional theory to check when the force-field proxy is reliable and when it is biased.
Shivang Agarwal, Rodrigo Wang, Kevin Ryczko et al.
In briefRecurrent In-Context Learning for Tabular Foundation Models: Replacing Distinct Layers with Looped Shared-Weight Blocks
SummaryThis paper asks whether the in-context learning stage of tabular foundation models needs a separate set of parameters for each layer or only sufficient depth. The authors propose Tab-Loop, which replaces the 12 untied ICL blocks of TabICLv2 with 4 tied blocks applied k times, leaving the rest of the model unchanged. At matched compute, Tab-Loop uses 62% fewer parameters while matching or slightly exceeding the untied baseline on held-out log loss, accuracy, and Elo benchmarks.
In briefCertify the Decision, Not Every Score: Exact Top-k Selection with Few Evaluations
SummaryThis paper introduces Certified Adaptive Selection (CAS), a method for finding the exact top-k robust algorithms or schedules without evaluating every candidate. It keeps guaranteed bounds on Expected Shortfall, the average loss in the worst scenarios, and stops once remaining uncertainty cannot change the selected set. Predictions only choose which outcome to evaluate next and never authorize the final decision.
In briefHow PFNs Depend on Feature–Label Dependence in Their Priors
SummaryThis paper studies prior-data fitted networks (PFNs), transformers pretrained on simulated data to approximate Bayesian predictions. It shows that common PFN priors are non-separable, meaning feature distributions and label-generating mechanisms are not independent, which violates standard theory. The authors characterize how this changes predictions, finding the difference vanishes under posterior concentration but can trade conditional label fit for feature fit under misspecification. They also present SemiTabICL, a PFN pretrained to use unlabeled data, which improves accuracy over a supervised baseline.
Julius Kobialka, Manuel H\"ulskamp, Thomas Nagler et al.
In briefRegularizing Average-Reward Q-Learning and Planning Toward a Mixing Policy
SummaryThis paper studies average-reward reinforcement learning, where the Bellman operator contracts in span only when the Markov chain mixes. The authors assume a reward-agnostic m-step mixing policy and regularize learned policies toward it to control contraction and exploration. They provide finite-time analysis of Q-learning variants with multiple regularizers and examine Q-planning. Numerical studies support the theoretical predictions.
