In briefWhat History Adds: Auditing Calendar Controls and Timing in University Trajectory Forecasts
SummaryThis paper audits whether historical summaries improve forecasts in a rule-based university trajectory simulator, comparing seventeen ridge models and a persistence baseline. A six-month history lowers six-step prediction error relative to current-state prediction, but its advantage over a length-only calendar control is statistically uncertain. The authors also report that 28.20% of six-step fatigue predictions fall outside the simulator's [5,90] support. The work is an evaluation audit rather than a new memory architecture.
In briefLearning to forecast across multiple spatiotemporal hierarchies
SummaryThis paper proposes a graph deep learning framework that treats variations across different aggregation hierarchies of the same time series as multi-task regularization. The authors argue that training on a single fixed hierarchy causes over-specialization, while joint learning across hierarchies improves forecasting accuracy. They also introduce a top-down operator that produces coherent forecasts with linear time and space complexity in the hierarchy size. Experiments show gains over single-structure training and competitiveness with current state-of-the-art methods.
In briefJevAgentHarness: Training Decision Models That Share Control with LLM Agents
SummaryThe paper notes that LLM agents call a large model at every step, making long tasks slow and costly, and that most calls are multiple-choice decisions. The authors build AgentDecisionBench to evaluate decision models on agent tasks and propose JevAgentHarness, which fits a confidence threshold per choice on the agent's own trajectories and lets the LLM decide below it. They also train AgentDecision-4B, an open 4B model, and report that on tasks like its training data it made agents 81.1-89.0% cheaper and up to 2.5 times faster without losing 3 or more points of success.
In briefJev-LiteMAS: Separating Decisions from Generation in Multi-Agent Workflows
SummaryThis paper presents Jev-LiteMAS, a multi-agent framework in which a lightweight decision model (Jev) selects the next expert and screens intermediate artifacts, while generative LLMs handle execution and corrective feedback. Across four benchmarks and two backbones, it reduces generative-LLM tokens by 12.7-54.6% compared with LAS. Accuracy stays within 1.34 percentage points, and latency decreases in seven of eight settings.
Kaixiang Wang, Zhaijiacheng Zhou, Jiong Lou et al.
In briefCLAD: Keypoint matching using global line correspondences
SummaryThis paper proposes CLAD, a keypoint matching method that describes each keypoint by its proximity to matched lines, adding global geometric context. Traditional local descriptors such as SIFT and ORB can be unreliable in scenes with repetitive or weak textures. The method uses HOME to detect linear structures. Evaluations on the TUM RGB-D SLAM dataset and outdoor video show geometrically consistent matches that complement local descriptors, and the authors report improved reliability in repetitive scenes with lightweight computation.
In briefFinding Effective Prompt Evolution Strategies for Stress and Emotion Prediction
SummaryThis paper compares six prompt evolution methods on social media stress and emotion data across two model sizes. It reports that GEPA and its extension combining instructions from two prompts (GEPA+merge) improved over initial prompts in several settings. The authors also identify five associated prompting strategies and document observations on task and evaluation limits.
SeungHyun Yi, Seungjun Yi, Sriram Sattiraju et al.
In briefMemCut: Testing whether LLM agents truly forget user secrets from long-term memory
SummaryThis paper introduces MemCut, a benchmark for testing whether the long-term memory of LLM agents truly forgets sensitive user information when requested. It contains 456 hand-designed secrets with annotated evidence trails, and measures recovery both through the system's retrieval and from surviving objects. Results show that removing only direct disclosures leaves much higher residual recovery, and that LLM-guided deletion still leaves 39-76% residual recovery, with limited candidate coverage identified as a bottleneck.
In briefMulti-Reward Hacking: Aggregation-Induced Drift and a Trust-Anchored Defense for LLM Training
SummaryThe paper studies LLM post-training that combines a primary reward, such as correctness, with auxiliary rewards, such as brevity. It shows that symmetric aggregation can raise the total reward while primary-task performance falls. The authors propose Trust-Anchored Gating (TAG), which selects an anchor channel and grants auxiliary credit only when the anchor is satisfied. Experiments on mathematics and program synthesis with Qwen and Llama models report that TAG mitigates this degradation.
WhyCharacterizes multi-reward hacking under symmetric aggregation and proposes anchored gating, backed by convergence proofs and Qwen/Llama experiments showing clear gains.
In briefBeyond Invariance: Factor-wise Selective Handling of Shifts in Multimodal Learning
SummaryThis paper addresses distribution shifts in multimodal learning by proposing FaShift, a framework that diagnoses latent representation factors by their distributional changes, intervention effects, and cross-environment stability. It separates non-predictive, stably predictive, and predictive but unstable factors, and selectively suppresses, preserves, or attenuates each one instead of enforcing uniform invariance. Experiments on MOSI, MOSEI, and CH-SIMS report the best in-distribution accuracy and F1 on all three datasets, plus out-of-distribution gains over the strongest competitors.
In briefEdge-AI Fault Detection for Optical Wireless Links
SummaryThis paper presents a self-diagnosing optical wireless receiver built on an ESP32-S3 microcontroller that runs a TinyML classifier on the device itself. It extracts channel features such as bit error rate, received signal voltage, SNR, and ambient light baseline while recovering 1000 bit/s OOK data. An INT8-quantized model using 3.32 kB of SRAM reaches 74.68% overall accuracy across four operating conditions. On-device tests indicate real-time, deterministic fault classification.
Sarthak Bhattarai, Sheshadri Shikhar Bandyopadhyay, Suprabh Dhakal et al.
In briefWhat If the Watts Differed? A Counterfactual Reasoning Benchmark for AI Data Center Construction
SummaryAI data centers lock in key choices such as power source, cooling method, and site before construction, and these determine their capital, water use, and carbon path. The authors introduce DCBench-440, with 200 generative questions over STRIPS planning worlds and 240 multiple-choice questions grounded in quoted site records and emission equations, to test whether LLMs track what each commitment rules out. Open-weight models scored 12.5-71.5% on generative questions and 58.3-92.5% on multiple choice. No model exceeded 4.0% exact match when asked to list every action available at a decision gate.
In briefSelf-guided LLM reasoning to diagnose errors in vehicle routing problems
SummaryVehicle Routing Problems are fundamental to many real-world applications, and recent work uses large language models to solve these NP-hard problems, usually assuming error-free inputs. This paper systematically tests whether LLMs can detect and correct errors in VRP instances, finding limitations with complex instances or errors that heavily increase reasoning demands. The authors propose a self-guided framework in which LLMs autonomously generate code that gives structured insights for error diagnosis. Experiments across diverse settings and LLMs show significantly improved error detection, reaching near-perfect accuracy.
In briefLearning to Reason with Selective Internal Feedback: A Reinforcement Learning Method Without Ground-Truth Answers
SummaryThis paper introduces SIF, a reinforcement learning method that improves language models using their internal confidence signals instead of ground-truth answers or external verifiers. SIF compares ranked token-level self-certainty profiles across windows and peers, and filters samples and groups to decide which responses update the policy. Across three backbone models and 12 math and science reasoning benchmarks, SIF's final checkpoints achieve higher average accuracy than the best-validation checkpoints of Intuitor and Vigor.
In briefPriorityFEC: Deadline-aware streaming error correction for passing cached model state between AI agents at the edge
SummaryThis paper addresses redundant computation in edge multi-agent LLM systems, where shared context already processed by one agent must be recomputed by the next. It proposes PriorityFEC, which places high-value KV-cache blocks earlier in transmission and uses feedback-adaptive streaming forward error correction to protect them before the time-to-first-token deadline. Evaluated on Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct across several datasets, it reports better task quality than compared transport schemes, with the largest gains under strict deadlines.
In briefHow Calibration Language Affects Block-Influence Pruning in Multilingual Language Models
SummaryThe paper studies whether Block Influence (BI), a metric used to prune Transformer blocks, transfers from English calibration to Indian languages such as Hindi, Marathi, and Tamil across three models. Late-layer BI is 20.7-55.5% higher for the Indian languages than for English, so English-only calibration misses late-layer activity. Averaging BI over equal English, Hindi, and Marathi budgets mitigates perplexity damage for some models and depths, but the effect is not universal, and on Llama it severely harms Tamil, so multilingual calibration must be validated per architecture and pruning budget.
In briefRDL-Bench: A Benchmark for Task-Aware Question Answering over Relational Databases
SummaryThis paper introduces RDL-Bench, a benchmark for evaluating natural-language question answering over real relational databases. It separates questions into reasoning-oriented ones, answered from the current database state, and prediction-oriented ones, which depend on future outcomes under temporal constraints. Each instance includes the question, task metadata, schema, temporal context, executable evidence or target-construction logic, and ground-truth answers. The benchmark supports both native solvers and task-aware modular systems where a router sends each question to a suitable execution module.
In briefPerception or Reasoning? Diagnosing RL Bottlenecks in Vision-Language Models
SummaryThis paper introduces PARA, a method that measures how much of each model's errors on a benchmark could be fixed by better visual perception versus a stronger reasoner. Across 4 vision-language models and 17 multimodal benchmarks, reinforcement learning gains are smaller when perceptual headroom is large and larger when reasoning headroom is large. The authors propose GRAD, which uses the best rollout in each group to directly guide the visual attention of weaker rollouts, and report consistent gains across two model scales and seven benchmarks. GRAD also complements existing RL methods.
In briefDPA-MoE: A Runtime Precision Control Framework for Mixture-of-Experts Inference
SummaryMixture-of-Experts models route each token to a few experts, but those experts run at a fixed precision. DPA-MoE encodes expert weights in chunks that can be decoded at different precisions and assigns chunks at inference time using live router gate scores. A single encoded checkpoint can therefore run at several weight budgets without retraining. Experiments on Qwen3-30B-A3B and DeepSeekMoE-16B report better KL divergence and downstream accuracy than uniform allocation at matched budgets, with modest throughput overhead.
In briefDoes Compression Imply Generalization? An MDL View
SummaryThe paper uses the minimum description length (MDL) principle to propose prequential regret, a compression signal computable from the training trajectory alone. In the nearly interpolating regime, its expectation upper-bounds the trained parameter's description length, yielding a high-probability generalization bound. Experiments on CIFAR-10 and CIFAR-100 show Pearson correlations above 0.9 with generalization error.
In briefMemorilla: A Latent Semantic Memory System for Frozen LLMs
SummaryThis paper presents Memorilla, a memory system that gives frozen large language models access to external knowledge. It compresses an entire document collection into a fixed number of query-dependent memory tokens. With only 16 memory tokens, it outperforms baselines such as Mem0 and MemoRAG on five of eight benchmarks. On FactKG it reaches 84.33% accuracy versus 80.43% for full-context inference, and it also improves an agent's success rate in a TextWorld environment.
Adibvafa Fallahpour, Parsa Idehpour, Vignesh Kothapalli et al.
In briefSame Target, Different Origins: Aligning Latent Rollout for Consistent PDE Emulation
SummaryNeural PDE solvers often predict multiple steps by autoregressive rollout, so errors can accumulate and make predictions at the same target time depend on their starting time and rollout depth. The paper proposes Align Latent Rollout (ALR), a regularization method that aligns the solver's internal latent representations across rollout origins, using zeroth-order state consistency and first-order dynamics consistency. ALR can be added to different neural PDE solvers without changing their architectures. Experiments on several PDE benchmarks and neural-operator backbones show improved prediction accuracy and reduced shared-target discrepancies.
In briefESG-Kor: A Korean Dataset for Extracting ESG Information and Its Practical Uses
SummaryThis paper presents ESG-Kor, a Korean dataset for automatically extracting Environmental, Social, and Governance (ESG) information. It contains 118,946 sentences taken from Korean companies' sustainability reports and manually labeled according to ESG evaluation agencies' rules. Several Korean pre-trained language models achieved significant classification performance, and the authors show potential use cases, including extending the model to small and medium enterprise documents and analyzing key ESG issues.
In briefSATOR: A backpropagation-free method for adapting 3D models at test time by routing tokens per sample
SummaryThis paper addresses test-time adaptation for 3D point clouds under distribution shift, without needing source data or backpropagation. SATOR computes token abnormality scores from attention and geometric graphs and uses Otsu-based separation to split retained and pruned tokens per sample, avoiding a search over candidate purge sizes. Keep-ratio-calibrated routing then chooses the pruned prediction when its penalized entropy is lower. The authors report state-of-the-art accuracy on ModelNet40-C, ScanObjectNN-C, and ShapeNet-C.
In briefRenoLayout: A unified framework for constraint-aware home floorplan renovation and furniture layout
SummaryThis paper presents RenoLayout, a framework that turns an existing floorplan and a set of design points into an executable Geometry-as-Code program, handling renovation and furniture layout together. The authors build a Design Map of 123 fine-grained design points and automatically derive design-point sets from before-and-after design pairs. They also release the RenoLayout Benchmark, which contains 200 real-world renovation cases with complete furniture layouts.
In briefCoDeFont: One-Shot Font Generation via Multi-Scale Factor-Specific Representation Learning
SummaryThis paper presents CoDeFont, a framework for one-shot font generation that synthesizes a target glyph by transferring style from a single reference glyph while keeping the character's structure. It extracts style and structural information separately across multiple spatial levels and uses a contrastive objective to separate style and content embeddings. The disentangled representations are injected progressively during decoding. The authors evaluate on structurally distinct Bengali and Latin scripts, on both seen and unseen font styles, and report consistent gains over existing state-of-the-art methods.
Rejoy Chakraborty, Prasun Roy, Saumik Bhattacharya et al.
In briefTwo sample sizes in weighted LLM ensembles: buffered ERM and minimax limits
SummaryThe paper studies how two sample sizes, the number of labeled questions Q and the number of generations N per model and question, affect learning the weights of a test-time LLM ensemble. Under a margin condition, it shows that empirical risk minimization with a calibrated margin buffer achieves excess risk of order Q^{-1/2}+N^{-β/2}, up to logarithmic factors. For a simplified two-model, two-answer submodel, it characterizes the minimax excess risk and identifies a floor of order N^{-β} that persists even with infinitely many questions.
In briefLTM: Latent Transition Memory for Robot Vision-Language-Action Models
SummaryRobots often need to remember interactions they can no longer see. This paper introduces Latent Transition Memory (LTM), which stores history as compact latent transitions between consecutive observations, implemented as LTM-π on π0.5. It reports 58.54% success on RoboMME and 80.5% on LIBERO-Mem, the best among baselines that use no external reasoning models or memory-specific annotations.
In briefL2GSSL: A label-efficient self-supervised method for learning from 3D medical CT scans
SummaryThis paper presents L2GSSL, a self-supervised framework that pretrains on unlabeled CT volumes by separating learning into intra-slice and inter-slice representation spaces. On LIDC-IDRI lung nodule malignancy classification, it reaches 0.9229 AUROC with 10% of labels and 0.9377 with 25%, outperforming the volume-level VICReg baseline. The authors report that its advantage is greatest when annotations are limited.
In briefIs the embedding table overpriced? A frozen 16 MB vocabulary for small and ternary language models
SummaryThis paper examines the word embedding table in small language models, which holds 168M of Gemma 3 270M's 268M parameters. The authors replace it with a frozen 16 MB table of 512-bit codes, derived once from a pretrained embedding via principal-component projection and sign binarisation, and compare it with a learned table on the same training data. The frozen-table model is better than or tied with the learned table on LAMBADA, HellaSwag and PIQA, while only ARC-Easy favours the learned table. With a ternary body, the model is a 46 MB file and runs several times faster on a laptop CPU than the float model in llama.cpp.
In briefIs Honesty the Best Policy? Demand Management with Probabilistic Service Options
SummaryThe paper studies service providers that display probabilities of predicted service levels at booking, where the true probabilities depend on the system state and customers cannot readily verify them. The authors model this as a sequential decision process, define truthful probabilities recursively, and introduce transparency constraints limiting distortion. They propose a reinforcement learning approximation and a sampling-based mixed-integer program to optimize disclosed information. A computational study on out-of-home delivery reveals a trade-off between individual-level transparency and aggregate customer utility.
In briefA graph neural network method for spotting low-quality, inconsistent documents
SummaryThis paper introduces a hierarchical graph convolutional network (HGCN) that detects internal inconsistencies in documents and mismatches between titles and bodies. The authors built an Inconsistency Dataset from published news and its metadata to train the model. The model reached 91.20% accuracy on this dataset and 92.00% on a public incongruity dataset, and an ablation study confirmed that metadata improves performance.
In briefFedSkillGit: Federated Skill Evolution Using Version-Control-Style Commits
SummaryThis paper proposes FedSkillGit, a framework that evolves reusable natural-language skills for LLM agents across distributed data owners without sharing raw data. Instead of uploading whole skill documents, each client submits a bounded set of rule patches and a privacy-filtered local experience summary, which the server deduplicates and merges into a candidate skill. The candidate is published only if clients' local validation scores improve. The authors report outperforming state-of-the-art baselines by up to 22% across three datasets and three model families, approaching centralized evolution.
In briefEmotion as Internal Signal: Attention Boosting for Affect Reasoning in LLMs
SummaryThis paper studies whether large language models can infer implicit emotional states from sparse conversational cues and whether they contain internal attention signals for routing such evidence. The authors propose Self-targeted Internal Attention Boosting (SIAB), which uses causal patching on synthetic minimal pairs to locate sparse affective attention heads and boosts the content tokens they attend to. On ECC and GoEmotions, SIAB improves prompt-matched baselines by up to 10.3 and 5.5 macro-F1 points.
In briefARGUS: An Agenda-Driven Framework for Reliable User Simulation
SummaryThis paper presents ARGUS, a neuro-symbolic framework for simulating users in interactive environments to evaluate and train agents. It tracks the user's goals with a dynamic agenda stack manipulated by symbolic operations, separating state tracking from language generation. The authors also distill ARGUS into compact models and introduce a rubric-based metric, Composite Behavioral Score (CBS), with evaluations on τ³-Bench and MultiWOZ.
In briefOpCredit: Operation-level policy learning for agentic knowledge graph question answering
SummaryThis paper addresses agentic knowledge graph question answering, where trajectory-level rewards cannot show which individual operations actually helped. The authors propose OpCredit, which combines branch outcome credit and frontier transition credit to supervise each operation's decision span during training. Inference uses only the learned policy and the standard KG interface. Experiments on ComplexWebQuestions, WebQSP, and GrailQA with three backbones report consistent performance gains.
In briefQuantum vs. Classical Machine Learning for Brain Disorder Classification
SummaryThis paper compares six quantum machine learning model families with classical models for classifying neuropsychiatric disorders from resting-state brain connectivity data, using identical features across four public cohorts totaling 4,308 subjects. Most quantum models performed comparably to their classical counterparts, but none consistently outperformed them. The study provides a feature-matched benchmark of 284 quantum-classical comparisons.
In briefOnline quantile omniprediction for proper losses, a method for combining forecasts accurate under many metrics at once
SummaryThis paper studies omniprediction, building a predictor that performs accurately under multiple evaluation metrics at the same time. Forecasts are expressed as quantiles over a discrete set of probability levels. The authors present sample-efficient online algorithms that remain accurate across all proper losses meeting mild regularity conditions. They demonstrate the approach by ensembling expert quantile forecasts of COVID-19 hospitalizations.
Hyunsuk Kim, Isaac Gibbs, Michael W. Mahoney et al.
In briefDimension-free PAC-Bayes generalization bounds for deep networks using Dirichlet processes
SummaryClassical PAC-Bayes bounds place priors on network parameters, so the bounds become vacuous as parameter count grows. This paper instead places a Dirichlet Process prior on the space of data distributions and derives parameter distributions through the learning algorithm. Using the data processing inequality, the resulting certificate has a closed form and does not depend on parameter count, width, or depth. On MNIST with models from 7,850 to 4,203,522 parameters, the certificate stays between 8.8% and 9.4%, while classical weight-space bounds exceed 940% and become vacuous.
In briefDoes on-policy distillation need to finish every generated response?
SummaryThe paper proposes Stochastic Truncated On-Policy Distillation (STOPD), which randomly stops student rollouts early and corrects the resulting sampling bias with blockwise inverse-probability weighting to cut computation. The authors prove the weighting preserves Vanilla OPD's gradient sum in expectation and derive one-step update-error bounds for SGD, Adam, and Muon. Across multiple model pairs, tasks, and optimizers, STOPD reduces generated tokens by 25-48% with accuracy comparable to Vanilla OPD and achieves 1.3-2.2x training speedups.
In briefSelecting and splitting attention heads to fine-tune language models with reinforcement learning
SummaryThis paper studies sparse and continual reinforcement learning fine-tuning of language models by ranking and partitioning attention heads. For sparse learning, heads are ranked by gradient norm to pick a small subset to train. For continual learning, heads are split across tasks to reduce interference and retain prior abilities. Experiments on math, instruction-following, and coding tasks show tuning higher-ranked heads often outperforms lower-ranked ones and approaches denser baselines.
In briefWhen Should Low-Light Enhancement Use Generation? Auditable Risk Control for Selective Generative Enhancement
SummaryThis paper addresses when generative low-light enhancers should override restoration networks, since generative methods add plausible texture but may hallucinate structure while restoration methods keep structure but may under-enhance. The authors propose VetoNet, which decides region by region whether to override a frozen base model and refuses the switch when evidence is insufficient. Experiments show statistically significant certified gains when expert complementarity is predictable, and conservative vetoing when routing evidence is weak or affected by cross-domain shift.
In briefSemantic latent alignment for flow matching using invertible bi-Lipschitz adapters
SummaryThe paper keeps a pretrained encoder and decoder fixed and adapts only the coordinates between them, using an invertible bi-Lipschitz adapter aligned offline to a semantic teacher and frozen during class-conditional flow matching. The bi-Lipschitz constraint limits how much alignment can distort the latent geometry. On two medical imaging datasets, constrained adaptation improves feature metrics, seed stability and counterfactual edit strength while staying competitive on downstream classification. Unconstrained adaptation, though still invertible, can become highly anisotropic and sensitive.
Felix N\"utzel, Mischa Dombrowski, Luca Hagen et al.
In briefVoteGEO-Bench: Measuring How Web Rewriting Affects Election Visibility in AI Answers
SummaryThis paper introduces VoteGEO-Bench, a benchmark covering the 2024 U.S. presidential and 2026 Japanese House of Representatives elections, to study how rewriting web content (generative engine optimization) changes election information in LLM responses. The authors analyze the sources generative engines cite and evaluate 11 rewriting strategies across four categories. They report that target-specified rewriting more often increases target exposure, and that improving page organization or strengthening a position through argumentation yields larger gains than adding evidence or changing wording.
WhyIntroduces a new election-focused GEO benchmark with 11 rewriting strategies across two countries, offering fairly solid evidence and notable societal impact on election information integrity.
In briefHybrid Decoding Combining Tagging and Generation for Product Attribute-Value Extraction
SummaryThis paper proposes a hybrid model that combines span tagging with open-ended generation to extract product attribute names and values. Frequent attributes with copyable values are extracted directly from the source by a tagging head, while a generative head handles rare attributes and non-copyable values. Integrated into vLLM inference, the hybrid decoder raises attribute coverage by 10% and extracts 63% more correct attribute-value pairs without precision loss, while reducing latency by 44%.
In briefDocPilot: an agent framework that helps small models answer questions about long, visual documents.
SummaryDocPilot is an agentic framework for answering questions about long multimodal documents by locating scattered evidence across pages and reasoning over text, images, and numbers. It combines page-level navigation, fine-grained evidence localization, local visual inspection, and code-based numerical reasoning. The authors also propose curriculum-based training to teach small language models these abilities, and report that DocPilot-4B performs competitively with much larger models on several benchmarks.
WhyProposes learnable document agent with curriculum training; gains across four benchmarks, and a 4B model competes with much larger models.
In briefCertEx: Exact GPU Image Search Using Certified Exclusion
SummaryThis paper presents CertEx, an exact k-nearest-neighbor search method for GPUs aimed at visual retrieval over image descriptors and embeddings. It stores a lower-bounding projection of each vector in half precision and uses a rounding certificate to exclude vectors that cannot be neighbors, computing exact distances only for the few remaining candidates. The authors prove that CertEx returns the exact kNN for every query and report higher throughput than the fastest cuVS search on 16 datasets with up to 100M vectors.
In briefAdaLoop: Reasoning by Repeatedly Running the Vision Encoder
SummaryMost vision-language models encode an image once with a static vision encoder, regardless of the question. This paper presents AdaLoop, a framework that adapts visual computation to each question by reusing the encoder's transformations over multiple iterations while keeping pretrained vision weights frozen. The authors also introduce VisRO, a procedural dataset and generation pipeline for training and evaluating vision-centric reasoning.
In briefCoViG: A Coarsened Vision Graph Network for Efficient Image Classification
SummaryThis paper presents CoViG, which combines graph coarsening with Vision Graph (ViG) models for image classification. It adds a MinCUT pooling step inside the Grapher component of intermediate ViG blocks to reduce the number of nodes in those blocks. Evaluated on ImageNet across several ViG architectures, CoViG reportedly matches the accuracy of the original ViG models while using about 51% of their FLOPs.
In briefZoomCap: Evidence-routed visual zooming to transfer recoverable visual information into captions
SummaryThe paper presents ZoomCap, a method that uses a draft caption to guide knowledge retrieval, text reading, gap recovery, and adaptive zooming on image regions, producing detailed captions for training vision-language models under a finite observation budget. It also introduces FactCycle, which scores atomic claims in captions against the source image and a reconstruction. The authors report human preference results and gains in continual pre-training and medical pre-training experiments.
WhyProposes evidence-routed visual zooming and fact-level evaluation, validated across several VLM pre-training transfer experiments.
In briefConstrained multi-agent reinforcement learning done in imagination, using a world model that predicts cost
SummaryThe paper introduces CA-MAMBA, which predicts cost in a multi-agent world model so that most constrained policy updates happen in imagination. It reports that on the median SMAC configuration, the stronger model-free baseline needs 10 to 20 times CA-MAMBA's interactions. On MAMuJoCo at equal budget, CA-MAMBA earns 2.8 to 12 times the baselines' reward.
In briefLoud and Clear: Improving How Retrieved Document Pages Are Fed to Vision-Language Models
SummaryThis paper finds that even when correct document pages are retrieved, vision-language generators underuse the visual evidence in them. It proposes Loud and Clear (LoC), which combines higher-resolution full pages with OCR from query-relevant regions, requiring no ground-truth annotations or extra training. LoC outperforms eight prior methods on SciEGQA, ViDoRe-v3, and MMLongBench-Doc, including the strongest baseline ColQwen3.
In briefMoney-Mule Detection as Relational Learning: A Study on the RBIH NFPC Dataset
SummaryThis paper proposes detecting money-mule accounts by scoring each account directly, treating the bank's relational database as a heterogeneous graph and using multi-hop neighbourhoods with a graph neural network. The model uses only evidence recorded before each account's flag cutoff. On a five-year RBIH retail-banking dataset, it recovers about nine in ten held-out mules at 83% precision and a 0.5% false-alarm rate, finding 115 more mules than the strongest tree ensemble given the same evidence.
In briefSpectrum Alone Is Not Enough: Random-Matrix Analysis of Data-Free Preconditioning for Private Learning
SummaryThe paper studies building KFAC Kronecker factors from synthetic probes instead of private data, using random-matrix theory to describe what such probes can reconstruct. It shows the error-signal factor G can be reconstructed without data, while data-induced spikes in the activation factor A cannot, and gives a threshold where this fails. Experiments find that even a probe with a nearly exact matched spectrum inflates residuals when its eigenvectors are random, so spectral matching is not the operative factor.
WhyRandom-matrix analysis of data-free probes gives a proven data-insensitivity bound and BBP threshold, validated across four modalities.
In briefPocketTSLM: a phone-run time-series language model that keeps health data on the device
SummaryThe paper presents PocketTSLM, a smartphone deployment of the soft-prompt OpenTSLM model that feeds encoded time-series windows as embeddings rather than as text. The authors report that, in their benchmark format, it uses about 12 times fewer time-series tokens than a text-serialized baseline and processes sleep EEG and 12-lead ECG clips. All inference runs on the device, and no health data is transmitted during inference.
Max Rosenblattl, Nikhil Krishnaswamy, Leon Nissen et al.
In briefVLA-Δ: Teaching Robot Models to Learn Physical State Changes
SummaryVision-language-action models apply knowledge from vision-language pretraining to robot control, but most map observations directly to actions without supervising intermediate physical state changes. This paper proposes VLA-Δ, which uses physical changes between current and future states as explicit learning targets, supported by a large-scale pretraining dataset with state-transition annotations. The model generates actions directly from current observations at inference time. It reports performance competitive with state-of-the-art methods on LIBERO and RoboTwin and outperforms π0.5 in real-robot experiments.
In briefVALCON: Testing and Improving Whether LLMs Judge Values Consistently Across Generating and Evaluating
SummaryCurrent evaluations of value alignment in large language models mainly check whether outputs comply with values, without testing consistency when the model switches between generating and evaluating. The authors introduce VALCON, which builds three benchmark suites by pairing a model's generations with its judgments under the same principles. Experiments on 28 LLMs show widespread cross-role inconsistency, where models with compliant outputs often judge their own generations differently. The paper also reports a role-dependent representational decoupling found through mechanistic probing and proposes a representation-level intervention that reduces these inconsistencies.
WhyIntroduces role-consistency as a new alignment dimension, validated across 28 LLMs with mechanistic probing and a mitigation intervention.
In briefDusting the Phantom: A Gaussian Allocation Method for Recovering Fast-Moving Dynamic Content
SummaryThis paper addresses a limitation of dynamic Gaussian representations for volumetric video, where small fast-moving objects and fine textures in highly dynamic regions can be lost. The authors argue that the bottleneck lies in where and when Gaussian primitives are allocated, not in representational capacity. They propose Seed Injection, which finds under-reconstructed dynamic regions and initializes new Gaussians at plausible 3D locations, and Temporal Biased Relocation, which uses time-resolved optimization signals to guide relocated Gaussians. Both keep the Gaussian budget fixed. Experiments on multiple dynamic-scene datasets report improved reconstruction quality, especially in highly dynamic regions.
In briefMaking medical device document assessment with LLMs repeatable at every stage
SummaryThe paper examines why large language models give different compliance assessments for the same medical device document request. Across fourteen commercial deployments, every one returned differing results for at least one request, and twelve changed the label. The authors built a pipeline that is deterministic at each stage, from text extraction through retrieval and inference, and got byte-identical outputs for blood glucose meter and infusion pump documents. They note that GPU model and inference engine scheduling settings also affect outputs and should be recorded.
In briefGeoPano: Geometrically Accurate 3D Reconstruction from a Single Panorama
SummaryGeoPano is a feed-forward framework that recovers a complete 360-degree point cloud from one panoramic image. It introduces a Pano-Attention mechanism that uses the panorama's spherical geometry to reduce inconsistencies between perspective views. A hybrid supervision strategy enforces both local detail and global structure. The authors report that it outperforms prior state-of-the-art methods in accuracy, geometric consistency, and visual fidelity.
In briefAgentic Interpretation (AgInt): Harness-Governed Evidence Construction for Software Agents
SummaryThe paper presents AgInt, a framework in which a model proposes typed evidence records that a deterministic harness admits, grades, and persists, then licenses answers under task-specific rules to produce an auditable certificate. It is evaluated on patch-acceptance benchmarks (SWE-bench-Live, DeepSWE) and fault-localization benchmarks (Defects4J, GitBug-Java). The authors report gains over the plain agent and Agentic Code Reasoning on several metrics, and a matched-template control to attribute which components drive each gain.
In briefPersonalized projected ascent for federated Stackelberg optimization
SummaryThis paper studies federated minimax problems in which the shared minimization variable is aggregated while each client keeps a personalized maximization variable constrained to a compact convex set. The authors show that projected ascent reaches the standard sublinear rate in the nonconvex-concave setting but hits a noise floor in the nonconvex-PL setting. They propose VR-SGDA, a variance-reduced method, which removes this floor and is supported by experiments on subpopulation robustness and NLP adversarial training.
M Yashwanth, Sharannya Ghosh, Saikiran Bulusu et al.
In briefCan Capability Enhancement Be Localized? Whether removing a few attention heads can improve task performance
SummaryThis paper studies Capability Enhancement Localization (CEL), asking whether a small set of attention heads in a Transformer language model can be removed to improve performance on a given task without further fine-tuning. The authors first find such sparse head combinations empirically across models from different families using a two-stage screening and joint-ablation routine. They then build a simplified multi-head softmax attention model with two conflicting tasks and show theoretically that gradient descent from small random initialization concentrates the conflicting features into particular heads, offering a training-based explanation for CEL.
In briefControlling and Injecting Bias to Aggregate Multiple AI Agents' Answers
SummaryThis paper examines how to decide whose answer to trust when multiple large language model agents disagree. The authors find that controlling non-semantic biases such as majority pressure and response style improves initial answer selection, while deliberately injecting bias produces decision instability that signals reliability. Building on this, they propose a test-time aggregation framework requiring no validation data, which reports higher average accuracy and token efficiency than competing non-oracle methods across three math reasoning benchmarks and four heterogeneous model combinations.
In briefReState: A training-free method that corrects object removal artifacts in few-step diffusion models
SummaryThis paper studies training-free object removal, where few-step distilled diffusion models often leave object-shaped residuals in the target region. The authors find that a structured prediction forms inside the mask at the first forward pass, while attention rerouting mainly acts in later steps. They propose ReState, which corrects the masked clean prediction toward background evidence from the input using a noise-level-aware weight. ReState achieves the best masked PSNR, SSIM, and LPIPS among training-free methods on two paired benchmarks and two few-step backbones.
In briefThe Density of Thought: Studying How Information Density in Reasoning Steps Predicts and Guides LLM Answers
SummaryThis paper examines whether the information density across chain-of-thought reasoning steps in large language models predicts whether answers are correct, and whether it can be adjusted at inference time. Experiments on six model scales, two training paradigms, and 60 GSM8K problems find that the relationship between density and correctness depends on the training method rather than model size. The authors also propose Entropy-Gated Density Routing, a self-supervised method that improves accuracy by 10.0 percentage points without external guidance.
Tanush Kalhan, Anupam Chattopadhyay, Sriram Ranga et al.
In briefMallea: Building, Steering, and Memorizing Virtual Worlds with a Coding Agent
SummaryThis paper presents Mallea, a system that builds virtual worlds from text prompts. A coding agent infers object interaction needs and plans object structure, properties, and rules, which govern how the world evolves. Users can steer the world through actions or text edits, and textures of whitebox-defined objects are stored and updated as those objects change. A renderer combines the current state with remembered appearance to produce observations.
In briefLICOA: Learning robot actions that steer projected light in mobile augmented reality
SummaryThis paper presents LICOA, a system where a robot decides how to move and where and how large to project content, based on a language instruction and camera views. It combines a driving chassis with a two-degree-of-freedom steerable mirror and builds a simulated environment with 150 tasks for training. In 480 real-world closed-loop trials across 12 tasks, the best adapted action model reached a 57.50% success rate, versus 8.33% when the same architecture was trained from scratch.
In briefDiviner: Human-governed AI agents that help find the root causes of large-scale infrastructure failures
SummaryThis paper presents Diviner, an agentic root-cause analysis system deployed at Meta for diagnosing failures in large distributed systems. It uses a four-stage lifecycle: encoding reviewed incidents and service-owner expertise into versioned skills, workflows, and a knowledge graph; executing investigations with evidence-linked hypotheses; evaluating saved trajectories against responder-reviewed outcomes; and adapting knowledge only after replay, quality checks, and owner approval.
Pradipta Ghosh, Gaurav Mitra, Shripad Nadgowda et al.
In briefTimeCap: A Reproducible Open Pipeline for Captioning Wearable Health Data
SummaryThis paper presents TimeCap, an open, deterministic pipeline that generates text captions directly from wearable time-series signals without using a language model. Applied to the OpenMHC dataset, it produced MHCap, a caption corpus of over 7 million captions across 19 HealthKit channels. The authors trained the OpenTSLM model on this corpus, and caption-augmented training beat a label-only baseline on 29 of 32 per-user binary evaluations across 8 health-outcome tasks.
Max Rosenblattl, Milana Gurbanova, Karl Aurel Deck et al.
In briefAuditing whether healthcare AI agents reliably escalate when workflow information is missing or conflicting
SummaryThe paper introduces a benchmark of 400 synthetic healthcare workflow episodes that tests whether language models choose to answer, ask for clarification, abstain, or escalate, scored under a pre-declared 5:1 cost for missed versus unnecessary escalations. Models reach 0.97 accuracy on clean inputs, yet uncertainty-prompted models miss 32–97% of required escalations. A transparent workflow-aware controller reduces missed escalations to 0.12–0.24, but a lexical stress test drops rule-detector recall from 1.00 to 0.00 on unseen paraphrases, so the controller's claims are limited to this benchmark.
Zihao Zheng (Washington University in St. Louis), Junyi Yao (Washington University in St. Louis), Baichuan Li (Southern Methodist University) et al.
In briefPRIME: Verifiable, evidence-grounded reasoning for LLM agents checking records
SummaryThe paper introduces PRIME, a method that fixes in advance which evidence routes can justify each conclusion before an LLM agent checks a record against requirements. A separate program logs check results and refuses conclusions the results do not support, and it directs or runs further checks for stronger conclusions. On 480 public-record tasks, a requirements-shown agent gave unjustified conclusions on 26.7% of tasks, while PRIME gave none. Its average score for justified conclusions rose from 30.1 to a maximum of 59.2 out of 100.
In briefTRACER: Trajectory-Aware Tool Retrieval for LLM Agents
SummaryAI agents choose tools from large registries, but showing the entire registry costs many tokens and can make selection worse as it grows. This paper presents TRACER, a retrieval front-end that indexes previously observed tool-call sequences in a trie and ranks likely next calls using a learned mix of empirical and contextual priors. On a mixed retail and telecom τ² environment, TRACER improved GPT-5.4's tool-selection accuracy by 4.7 points while cutting input tokens per step by 80%, exposing only 5 of 128 tools.
In briefTRACER: A multimodal framework that combines hospital video and clinical records to track patient behavior over time
SummaryThe paper presents TRACER, a training-free framework that turns continuous ICU video into minute-level action records and combines them with clinical records into hourly reports. An LLM agent uses tools to query these records, inspect video, and retrieve clinical data. Across 280 patients and about 60,000 camera-hours, patients with a positive delirium assessment showed significantly less purposeful activity than those without (p < 0.0001).
Zane Durante, Paul Tang, Keshigeyan Chandrasegaran et al.
In briefRethinking Learning from Label Proportions via Moment Matching
SummaryThis paper studies learning from label proportions (LLP), where training data are grouped into bags and only bag-level label proportions are observed. Many existing methods assume random bag construction, which the authors find leads to suboptimal performance in non-random settings. Under a more realistic assumption of i.i.d. bags with conditionally independent labels, the authors show label counts follow a Poisson multinomial distribution. They propose LLP-MM, which uses multi-order factorial moments as training objectives, and report effectiveness and computational efficiency on benchmark datasets.
WhyDerives a Poisson multinomial distribution under realistic bag assumptions and uses factorial moments as objectives, with broad benchmark experiments across bag constructions.
In briefChecking whether sparse autoencoder features combine additively in model behavior
SummaryThe paper studies whether the effects of sparse autoencoder (SAE) features on model behavior add up simply. It introduces SAE Additivity Residuals (SAR), which measure behavioral variation left unexplained above a chosen interaction order, and compares learned SAE coordinates with random rotations of the same subspace. Learned coordinates show lower first-order residuals across all 18 settings, but pairwise interaction terms still improve held-out ordering error, indicating that faithful joint explanations require interactions beyond individual feature effects.
WhyProposes interaction-order additivity residuals, validated against random rotations across 18 settings with consistent results, offering moderate influence on SAE interpretability research.
In briefFractional Power-of-Two Quantization for Multiplier-Free LLM Inference
SummaryThis paper proposes Fractional PoT (FPoT), a base-√2 weight quantization grid that enables 4-bit LLM weights while keeping a shift-add integer datapath. On Llama-3-8B with 4-bit weights and activations, it improves zero-shot commonsense reasoning by 8.72% over the PoT baseline, staying close to the uniform integer baseline (0.4 percentage-point gap). The authors state the method can be combined with existing LLM post-training quantization approaches.
In briefSeparating a language model's judgment from its expression using a label-free subspace
SummaryThe paper studies why changing instructions for the same input can make a language model's final answers less reliable. The authors find that differences in hidden states caused by instructions concentrate in a low-dimensional expression subspace, which they identify without benchmark labels and remove from a linear predictor to isolate the judgment. Across correctness, entailment, and safety tasks, the isolated judgment stays reliable under unseen instructions, raising AUROC for entailment and safety under negated questions from 0.504 to 0.825 and from 0.480 to 0.891.
In briefLLM-based weak labeling for classifying cybersecurity tweets using a teacher-student setup
SummaryThe paper addresses the shortage of expert-labeled data for classifying cybersecurity discussions on social media. It proposes a teacher-student weak supervision framework in which a large language model generates labels for unlabeled tweets to train cybersecurity-specific pre-trained language models. Models trained only on these LLM-generated labels reportedly achieve F1 scores above 0.92, within 3 percentage points of fully human-supervised models.
Ferry Afit Kurniawan, Kazushi Okamoto, Koki Karube et al.
In briefHow AI chatbots respond after users refuse crisis referrals for suicidal thoughts
SummaryThis cross-sectional audit tested AI chatbots using 2,550 scripted multi-turn conversations in which simulated users disclosed suicidal thoughts and repeatedly refused suggested outside help. Nearly all chatbots offered a referral initially, but how long they kept pressing it varied widely, and newer versions persisted less than older ones. Disclosures of intent were linked to more sustained refusals and higher odds of renewed referral after the topic changed. The authors argue that safety evaluations should test what happens after a referral is refused, not only whether one is offered.
WhyAudits 2,550 multi-turn conversations after refused referrals, quantifying persistence differences and newer versions persisting less.
In briefDo VLA Models Really Follow "How" Instructions? A Diagnostic Study
SummaryThis paper examines whether vision-language-action (VLA) models can follow How-level instructions, which describe how to perform a task rather than just what the task is. The authors find that such instructions do not help current VLAs and can lower task success. Training with segment-level enriched instructions makes actions more responsive to them, though task success does not consistently improve. They also introduce SIMPLER-IN, a benchmark where success depends on conditions given only in the instruction, showing the gains do not extend to demanding or out-of-distribution instructions.
In briefFocusOPD: structure-guided on-policy distillation that aligns deep, early-denoising features in diffusion models.
SummaryThis paper studies multi-teacher on-policy distillation for rectified-flow diffusion transformers, where a student's internal features may not match its teachers' even when outputs do. The authors find the mismatch is concentrated in deeper Transformer blocks during early denoising and persists after standard OPD distillation. They propose FocusOPD, which adds token-wise directional feature alignment restricted to those deep layers and early steps. It improves results over DiffusionOPD on GenEval, PickScore, and LAION-Aesthetic.
In briefFGSS-BO: Feedback-Guided Hypothesis Management for LLM-Based Bayesian Optimization
SummaryThis paper proposes FGSS, a structured interface that links language-model hypotheses in Bayesian optimization to observed evidence, executable parameter changes, and comparative support updates. An LLM proposes and revises hypotheses, while a deterministic curator validates them and tracks test records. Built on LLAMBO, FGSS-BO improves search performance over the matched baseline on two retrospective tasks, with gains that vary across six hyperparameter-optimization tasks.
In briefSystem1-Arena: Which Decisions Can Be Left to Fast, Single-Pass AI Models?
SummaryThis paper introduces System1-Arena, a benchmark that tests whether fast single-pass models, which output option probabilities directly, can handle discrete decisions as well as LLMs that reason step by step. It covers 92k questions across 105 tasks and 138 closed-loop environments, and evaluates 30 models. The largest System-1 models match or exceed fast closed LLMs on non-agentic accuracy with better calibration at one-quarter the latency. In closed-loop settings, large open LLMs succeed most often, while small System-1 models rarely do.
In briefA decentralized multi-agent coordination method that ties task delegation to clear responsibility.
SummaryThis paper proposes RCM, a coordination mechanism in which agents use commitments to link task delegation with explicit responsibilities. Agents keep their upstream responsibilities during recursive delegation while locally orchestrating downstream commitments. They establish, manage, and end commitments autonomously through a responsibility-oriented communication protocol (RCP), so global coordination emerges without central control. Experiments on reasoning, coding, and executable agent environments report gains over centralized and decentralized baselines in task performance and responsibility, and improved failure and harm detection without task-specific training.
Thi-Nhung Nguyen, Viet-Thanh Pham, Linhao Luo et al.
In briefRecurring Feature Shifts in Sparse Autoencoders and Their Effect on Downstream Decisions
SummarySparse autoencoders (SAEs) split language model activations into sparse features, and rewording prompts can shift activity among them. The paper introduces the redistribution sensitivity score (RSS), which finds large feature changes that recur in a consistent direction, and tests them on variants not used for selection. Across 27 pretrained SAEs from eight language models, the tested prompt transformations showed a reproducible redistribution pattern. The authors separate representation change from decision failure, and report recurrent failure AUROCs of 0.935 and 0.958 on two frozen safety readouts.
In briefA climate-aware satellite image dataset for adapting land-cover segmentation across regions
SummaryThis paper introduces CASID, a dataset of 980 large satellite images from 30 regions across Asia, covering over 24,500 square kilometers. The images come from four climate zones, each forming a separate domain, making it the first climate-aware land-cover segmentation dataset with multiple domains. The authors analyze differences among the climate zones and evaluate recent segmentation and unsupervised domain adaptation methods to provide baseline results.
In briefCompatibility Matters: Token-Level Teacher Selection for On-Policy Distillation
SummaryThis paper proposes selecting teacher models token by token during on-policy distillation (OPD), rather than using a single teacher per prompt. The authors theoretically analyze what makes a teacher and student compatible and derive empirical heuristics for teacher selection. Experiments in math, coding, kernel generation, and visual reasoning report up to about 10 percentage points higher pass@k than vanilla OPD and the teacher experts on some tasks.
In briefUpdate-Matched Unlearning: Removing Knowledge from LLMs Without Losing Learning Ability
SummaryThis paper studies how to make large language models forget selected knowledge while still resisting relearning of that knowledge and keeping their ability to learn new tasks. The authors propose Update-Matched Unlearning (UMU), which applies the same simulated relearning update to a defended model and to an undefended reference model, and uses new-task learning to guide model selection. Experiments on TOFU and FaithUn with Llama and Qwen models show that UMU improves relearning resistance and new-task accuracy across three base unlearning methods.
In briefWhat to Remove: Assessing the Reliability and Costs of Decontaminating RLVR Training Data
SummaryThe authors audit 16 open math training corpora for reinforcement learning with verifiable rewards, comparing about 1.67 million items against six benchmarks with 1,949 questions. They test eight cleaning policies using exact, lexical, and semantic matching, validated by blind dual human review. Most semantic thresholds failed to reach the 90% conservative precision target on calibration data, while a judge-adjudicated policy showed higher precision and recall on held-out corpora. The study measures removal reliability and data loss, not just overlap counts, and releases its labels and decisions.
In briefKBS: A Korean Benchmark Suite for Testing Large AI Models
SummaryThis paper introduces KBS, a Korean benchmark suite with six tests (Ko-ARC, Ko-GSM8K, Ko-EQ-Bench, Ko-WinoGrande, Ko-LAMBADA, Ko-IFEval) covering 9,396 items across six capabilities. It builds the tests through review, localization, direct Korean writing, and redesign rather than translation alone. The authors evaluated Korean-developed, multilingual open, and closed API models under protocols accounting for access, prompt format, and scoring differences. They find Korean benchmark scores depend on task construction, scoring rules, prompting, and post-training, so results should report both scores and the settings used.
In briefLagrangian-Eulerian graph neural ODEs for learning and forecasting multiscale epidemic dynamics
SummaryThis paper introduces LE-EpiGNODE, a model for medium- to long-range epidemic forecasting. It replaces generic message passing with mechanism-guided neural operators: age-contact data drives Lagrangian infection generation, while mobility drives conservative Eulerian population transport within a structure-preserving universal differential equation. Evaluated on 14 synthetic mechanisms and three COVID-19 datasets, it achieved the lowest WMAPE at the longest horizon on all three real datasets, reducing error by 17.7-40.0% relative to the best baseline.
In briefRecSoup: An Efficient Model Merging Method for Controllable Multi-Objective Recommendation
SummaryRecommender systems must often balance conflicting objectives. This paper proposes RecSoup, a lightweight two-stage approach that merges objective-specific models, first pruning parameters to reduce cross-objective interference and then adaptively merging them via a learnable preference-to-coefficient mapping. This allows trade-offs to be adjusted at inference time without joint retraining. The authors report improved Pareto front quality and controllability on multiple benchmarks, with inference latency 165 to 228 times faster than parameter generation approaches.
WhyApplies model merging to controllable multi-objective recommendation with pruning and learnable preference mapping; multi-benchmark experiments show Pareto quality and latency gains.
In briefAutoOnboarder: Automatically Adapting LLM Agents to a Task
SummaryThe paper notes that agentic large language models can lose reasoning ability when placed inside a harness, the software that mediates their observations and actions. It introduces AutoOnboarder, a system in which a supervising coding agent, the Onboarder, iteratively adjusts the model, prompt, and harness code to improve performance and reduce cost. On genomic-sequencing prior authorization form submission, the onboarded agents matched or exceeded a Gemini 3.5 Flash baseline at up to 29 times lower cost, though results varied widely and the agents remain imperfect.
In briefSeparating Depression from Cognitive Decline in Elderly Speech Using Concurrent Clinical Tests
SummaryThe study collected Korean speech from 89 elderly people with mild cognitive impairment, with depression (SGDS) and cognitive (MMSE) assessed at the same time. After controlling for cognitive function, only formant features were linked to depression, while commonly used F0 features showed no signal. A subset of just 12 F1+F2 features reached a UAR of 0.760, outperforming the full eGeMAPS baseline and self-supervised representations.
In briefDLOSAC: Joint Spatial-Geometric Image Estimation via Delaunay-Guided Local Optimization RANSAC
SummaryThis paper addresses image geometric estimation from noisy matching points, a task usually handled by RANSAC with local optimization. The proposed DLOSAC builds a Delaunay triangulation over candidate inliers and jointly uses edge-based geometric consistency, spatial distribution regularization derived from triangle areas and edge attributes, and point-level geometric constraints. The authors report state-of-the-art results on six public real-world datasets covering three types of geometric estimation tasks.
In briefRubric scores ignore judgment uncertainty; a conformal weighting fix for LLM evaluation
SummaryThe paper notes that rubric-based evaluation of LLM outputs aggregates binary per-criterion judgments uniformly, ignoring that some judgments are more certain than others. The authors propose URW, a training-free method that measures per-criterion uncertainty via conformal calibration over paraphrased judgment templates and converts it into weights. Experiments report improved pairwise preference accuracy and stronger downstream reinforcement learning alignment.
In briefSpaceCache: Reusing Intermediate Features to Speed Up Point-Cloud Segmentation
SummaryConsecutive point-cloud frames in autonomous driving and robotics largely overlap, yet existing segmentation networks process each frame from scratch. The authors propose SpaceCache, a training-free feature cache that reuses intermediate transformer features for previously observed regions and computes only the remaining tokens. On nuScenes, Waymo, and ScanNet, with the Utonia and LitePT-B backbones, it reaches up to 1.65x speedup with under 1 point of mIoU drop and outperforms token-merging methods.
In briefThinking makes reasoning models read words like "some" and "or" literally
SummaryThis paper examines how extra thinking changes the way reasoning models interpret everyday inferences. With model weights fixed and only thinking length varied, the authors find that thinking cancels the implicatures of "some" (not all) and "or" (not both) in most tested open models and GPT-5. The reasoning traces show models first consider the enriched reading and later cancel it.
In briefWho Can Cancel an Agent's Promise? Standing in Multi-Party Chat
SummaryThe paper asks who may cancel a commitment an AI assistant made to a team member in a multi-party chat. It defines "standing": only the creditor or someone they or the operator policy authorizes can release a promise, regardless of rank. Testing six models, it finds many override released promises or drop them on a peer's word. A self-kept commitment ledger and a short release policy markedly improve accuracy.
WhyFrames commitment release authority as a new problem, separates two error types, and validates ledger and policy fixes across six models.
In briefUltraDiff: Transferring High-Fidelity Priors to Compressed Latent Spaces for High-Resolution Image Generation
SummaryThis paper addresses the trade-off between computational cost and fidelity in Latent Diffusion Models, where stronger latent compression (f32) is efficient but loses information. It proposes UltraDiff, which combines joint latent diffusion learning with marginal score distillation to carry f8 priors into the f32 space. Across 1K to 4K settings, the method approaches f8 quality while keeping f32 efficiency, generating a 2K image in 6.2 seconds with 20 steps on a single A100 GPU, over 40 times faster than the f8 baseline.
